Evolutionary digital twin framework for optimal aminoglycoside dosing in neonates with suspected sepsis

Michela Prunella; Chiara Romano; Alessandro Borri; Nicola Altini; Maria Domenica Di Benedetto; Pieter Annaert; Karel Allegaert; Anne Smits; Vitoantonio Bevilacqua

doi:10.1038/s41746-026-02558-w

. 2026 Mar 21;9:367. doi: 10.1038/s41746-026-02558-w

Evolutionary digital twin framework for optimal aminoglycoside dosing in neonates with suspected sepsis

Michela Prunella ^1,^#, Chiara Romano ^2,^3,^✉,^#, Alessandro Borri ^2,^3,^✉, Nicola Altini ¹, Maria Domenica Di Benedetto ², Pieter Annaert ^4,⁵, Karel Allegaert ^4,^6,⁷, Anne Smits ^6,⁸, Vitoantonio Bevilacqua ¹

PMCID: PMC13168521 PMID: 41862656

Abstract

Aminoglycoside dosing in suspected neonatal sepsis remains difficult due to highly variable pharmacokinetics driven by marked physiological diversity, from extremely preterm to term neonates, and further complicated by acute kidney injury, perinatal asphyxia, and concomitant interventions. We developed multiscale medical digital twins combining a physiologically-based pharmacokinetic model with an eco-evolutionary pharmacodynamic module capturing drug-modulated bacterial growth and resistance. Glomerular filtration rate is continuously updated using a long short-term memory neural network trained on real-world data. Calibrated on 1634 neonates, the framework enables in silico optimization of full-course antibiotic therapy through real and virtual cohorts, balancing efficacy and safety while accounting for resistance-driven changes in the minimum inhibitory concentration (MIC). Nonlinear optimal control achieved bacteriostatic exposure across all digital-twin neonates, with safety preserved in most cases at higher MICs. Model predictive control further reduced bacterial rebound during late therapy. This framework supports evolution-aware precision dosing of renally cleared antibiotics in vulnerable neonatal populations.

Subject terms: Computational biology and bioinformatics, Microbiology

Introduction

The neonatal population is particularly vulnerable to the transition from local infection to systemic inflammation, such as sepsis, due to immature host-defense mechanisms. The pathophysiology of inflammatory responses in newborns differs markedly from that observed in adults and older children, and is further affected by maturational factors, such as gestational age (GA) and postnatal age (PNA)¹. Sepsis and other serious infections, occurring either within 72 h of birth (early-onset disease) or afterwards (late-onset disease), represent a major cause of neonatal mortality worldwide. Recent estimates from the World Health Organization (WHO) indicate that sepsis accounts for ~1.3–3.9 million cases and 400,000–700,000 deaths annually², with the highest incidence observed among very preterm and low-birth-weight infants³, particularly in low- and middle-income countries.

Since most neonatal sepsis is bacterial, with pathogens identified as Gram-positive or Gram-negative in blood cultures, effective treatment depends on selecting the appropriate antibiotic and achieving sufficient exposure to suppress bacterial growth. This requirement is typically quantified through the minimum inhibitory concentration (MIC), which represents the antibiotic concentration that inhibits the visible growth of bacteria strains in monoculture experiments⁴. However, in clinical settings, bacteriostatic effects depend not only on achieving the appropriate antibiotic concentration, but also on the time to administration⁵ and, for several antibiotic classes, on sufficient cumulative drug exposure over time⁶. Subinhibitory antibiotic exposure may decrease the susceptibility of important pathogens⁷, whereas higher doses for prolonged time can deplete commensal microbiota and contribute to the long-term development of antibiotic resistance^1,8. At the same time, inadequate dosing can compromise clinical treatment efficacy, leading to treatment failure and adverse outcomes, particularly in neonates^9–11. While recent regulatory frameworks, such as the 2022 FDA guidance on neonatal clinical pharmacology¹², have advanced the standards for neonatal dose selection and study design, residual limitations in the reliability of conventional dosing approaches persist. These limitations are primarily driven by the marked physiological heterogeneity of the neonatal population, frequent comorbidities, and the continued scarcity of adequately powered neonatal trials for many neonatal-specific conditions¹³. Although the FDA guidance cautions that extrapolation of effectiveness from adults is rarely feasible for neonatal-specific conditions, the scarcity of adequately powered neonatal studies forces a continued reliance on adult-derived PK/PD targets for many antibiotics, including aminoglycosides. In this context, Antibiotic Stewardship Programs (ASPs) have expanded in response to concerns about antibiotic misuse and the emergence of antimicrobial resistance¹⁴, aiming to ensure the continuous and judicious management of anti-infective therapies within clinical settings. In neonatal care, optimization of antibiotic therapy is therefore essential to minimize long-term risks, preserve immune system development, and reduce unnecessary healthcare costs¹⁵, with population-level relevance reflected by the WHO estimates indicating that up to 84% of neonatal deaths due to infections are potentially preventable through early diagnosis and timely, appropriate clinical management².

Evidence reports that the duration of treatment may vary from a minimum of 7 days to a maximum of 14 days for culture-proven, uncomplicated neonatal septicemia¹⁶. Within this treatment duration window, balancing pharmacokinetics-pharmacodynamics (PK-PD) efficacy against nephrotoxicity and resistance selection remains a major challenge.

In neonates, body composition, organ size and ongoing maturation change dramatically during postnatal life and, in combination with GA, influence not only the PK but also the PD of antibiotics. Accurate prediction of Therapeutic Drug Monitoring (TDM) concentrations in relation to developmental maturation, co-medications, and pathological conditions provides a foundation for rational dose and schedule optimization¹⁵. Conventional population pharmacokinetic (pop-PK) approaches, including simple compartmental¹⁷ and semi-mechanistic¹⁸ models derived from TDM concentration-time data, have demonstrated robust performance for exposure prediction and PK/PD target attainment. However, these approaches are primarily formulated to describe inter-individual and intra-individual variability statistically and are not designed to explicitly simulate the continuous, time-evolving physiological state of each patient. Recent approaches integrate body pathophysiology into ADME modeling, where compartments represent anatomically distinct and functionally homogeneous virtual units, such as organs or tissues. The level of physiological granularity governs parameter identifiability and bridges empirical pop-PK with physiologically-based PK (PBPK) or mechanistic models¹⁹. Longitudinal TDM data, combined with loop-based architectures such as long short-term memory (LSTM) neural networks, can then be used to estimate time-varying parameters in PBPK models²⁰.

Among the antibiotics commonly used for treatment of suspected neonatal sepsis, aminoglycosides remain widely prescribed despite the availability of newer antimicrobial classes. Gentamicin and amikacin, in particular, continue to represent key therapeutic options for the management of serious neonatal infections.

Amikacin is an aminoglycoside antibiotic effective against a wide spectrum of aerobic Gram-negative bacteria, including gentamicin-resistant strains, and is stable to most inactivating enzymes²¹. With negligible plasma protein binding, amikacin is predominantly eliminated via glomerular filtration, with minimal hepatic contribution. It exhibits concentration-dependent bactericidal activity, achieving optimal efficacy when the peak concentration ( $C_{\max}$ ) exceeds the MIC by a ratio greater than 8:1, while maintaining low trough levels ( $C_{\min}$ ) to minimize the risk of toxicity during therapy²². Other PK-PD targets include the area under the drug concentration curve (AUC) normalized by MIC (AUC/MIC), and the fraction of time the concentration is greater than MIC during a dose interval²². Although MIC-based PK-PD indices are useful for characterizing dose-response relationships, they rely on static, noise-prone in-vitro measurements that cannot capture the evolutionary pressures driving bacterial adaptation and the resulting inter- and intra-patient variability observed in vivo²³. PK model-informed dosing has guided antibiotic optimization for aminoglycosides²⁴, including neonates with perinatal asphyxia treated with therapeutic hypothermia (PATH)^25,26, and for other classes such as glycopeptides^27,28 and β-lactams^29,30. However, static PK-PD targets assume fixed pathophysiological characteristics. Moreover, existing approaches remain confined to initial dose design and short-term (24–48 h) simulations, without a unified framework to optimize long-term clinical outcomes throughout the Neonatal Intensive Care Unit (NICU) stay^29,31. In neonatal intensive care, PK is further complicated by comorbidities such as perinatal asphyxia, acute kidney injury (AKI), and patent ductus arteriosus (PDA), as well as concomitant interventions (e.g., PATH) and co-administration of drugs (e.g., ibuprofen, inotropes), which profoundly affect renal clearance, aminoglycoside exposure and bacterial adaptation³².

From an ecological perspective, both the magnitude and the temporal profile of antibiotic exposure define the selective environment, generating variability that serves as a substrate for bacterial adaptation. Resistance mechanisms, such as the expression of antibiotic-inactivating enzymes or the emergence of phenotypic variants that transiently reduce susceptibility, often come with a fitness cost, meaning that resistant bacteria typically grow more slowly or are less competitive than susceptible counterparts in the absence of antibiotics³³. The cumulative result of this selective pressure is the progressive acquisition of resistance traits that diminish drug efficacy, typically manifested as an increase in the MIC³⁴. Adaptive therapy paradigms have recently been proposed to exploit these eco-evolutionary interactions, maintaining controlled coexistence between sensitive and resistant subpopulations to delay full resistance fixation³⁵. Building on these principles, control-oriented modeling approaches treat pathogens as evolving populations to rationally design dosing strategies that both suppress total burden and modulate the emergence of resistance. Such frameworks typically integrate eco-evolutionary ordinary differential equations (ODEs) with multi-objective optimization to minimize pathogen load, delay resistance, and respect pharmacological safety constraints³⁶. Incorporating such evolutionary feedbacks through combination or cycling strategies may further enhance resistance-aware dosing design, yet their implementation in neonatal care remains largely unexplored.

Motivated by this unmet clinical need, we propose a comprehensive framework, illustrated in Fig. 1, that integrates PBPK modeling with an eco-evolutionary PD formulation for the optimization of neonatal antibiotic therapy. The framework is developed at the level of the aminoglycoside class and, in this study, is instantiated, calibrated, and validated using amikacin as a representative case study.

The proposed framework accounts for age-dependent renal maturation, therapy-induced bacterial selection, and the dynamic interplay between PK, bacterial adaptation, and ecological competition. A key component is an LSTM-based network for sequence-to-one regression, iteratively forecasting next-day glomerular filtration rate (eGFR), a critical PBPK parameter influencing amikacin disposition.

The calibrated PBPK-PD model was validated against measured amikacin concentrations obtained through TDM, demonstrating accurate predictive performance. In addition, the simulated bacterial dynamics during the NICU-administered therapy significantly stratified patients in accordance with the blood-culture-proven infection outcomes. This enabled the construction of a digital twin (DT) for each real neonate and the extrapolation of feasible parameterizations to generate a comprehensive virtual cohort. In this virtual cohort, baseline MIC values were sampled from clinical distributions and dynamically evolved during treatment in accordance with emerging resistance phenotypes³⁷. The coupled PBPK-PD system feeds a therapy optimization algorithm designed to operate under patient-specific physiological conditions, including clinical interventions such as AKI or PATH, to deliver individualized dosing strategies that systematically balance therapeutic efficacy against toxicity, dynamically adapting to evolving PK profiles and bacterial ecology. Specifically, we examined two control strategies: single-cycle optimization, which updates dosing once per cycle, and nonlinear Model Predictive Control (MPC)³⁸, which repeatedly re-optimizes therapy to handle phenomena such as late-cycle bacterial rebound. The resulting in silico feasible trajectories enable a comprehensive comparison between standard and DT-derived therapies, providing insights into bacterial outcomes while reflecting intra- and inter-patient variability³⁹.

To our knowledge, no existing neonatal dosing framework integrates these components nor provides a computational tool capable of adapting therapy in real time. Addressing this need, our approach demonstrates how interpretable, physiology-based models can advance precision antibiotic therapy in neonates, a clinically challenging and still underrepresented population²⁸.

Results

In this section, we describe how the mechanistic and data-driven components of the integrated PBPK-PD framework jointly construct patient-specific DTs and enable its application to optimize antibiotic therapy. Specifically, a four-compartment PBPK model predicts antibiotic exposure, while an eco-evolutionary PD module is calibrated to link this exposure to adaptive bacterial responses. The PBPK model was calibrated using real patient data, supplemented with physiologically based equations from the literature. This layer is complemented by a data-driven model for neonatal sCr prediction across cohorts, which accurately captures maturational renal function and consistently outperforms existing approaches. Taken together, these elements enable the full PBPK-PD system to reproduce observed amikacin PK and blood-culture positivity patterns in real neonates. The system is then used to generate a virtual cohort that recapitulates key maturational and pathophysiological trajectories, providing a physiologically coherent substrate for in silico exploration. Within this validated physiological and ecological landscape, the optimization module yields individualized dosing strategies that meet PK-PD objectives, satisfy safety constraints, and maintain control over resistant bacterial subpopulations. Its nonlinear MPC extension further enhances robustness by counteracting late-cycle bacterial rebound. Finally, the clinical translatability of the evolutionary DT optimizer is evaluated in real-world neonatal cohorts, demonstrating how DT-derived dosing improves therapeutic performance compared with standard clinical treatment.

A novel eco-evolutionary PD model links antibiotic exposure to adaptive bacterial response

We develop an eco-evolutionary PD model grounded in Darwinian dynamics to describe competition between susceptible (N_S) and resistant (N_R) bacterial subpopulations under antibiotic pressure. Eco-evolutionary frameworks underlying this approach were first explored in the field of tumor therapy optimization⁴⁰. Throughout the manuscript, the time dependence of state variables and parameters is omitted for notational simplicity, except in the control formulation where temporal evolution is explicitly considered. Each subpopulation i ∈ {S, R} is characterized by a per-capita fitness function f_i(C_b, N_S, N_R, u_i) that accounts for logistic growth, inter- and intra-strain competition, and antibiotic killing, given by:

f_{S} (C_{b}, N_{S}, N_{R}, u_{S}) = r_{S} e^{- c u_{S}} (1 - \frac{a_{S S} N_{S} + a_{S R} N_{R}}{K}) - \frac{E_{\max} C_{b}}{E C 5 0_{S_{0}} e^{k_{u} u_{S}}} + C_{b} - d,

f_{R} (C_{b}, N_{S}, N_{R}, u_{R}) = r_{R} e^{- c u_{R}} (1 - \frac{a_{R R} N_{R} + a_{R S} N_{S}}{K}) - \frac{E_{\max} C_{b}}{E C 5 0_{R_{0}} e^{k_{u} u_{R}} + C_{b}} - d .

Phenotypic resistance is introduced as a key eco-evolutionary component through a strain-specific adaptive trait u_i. By definition, the susceptible strain does not express resistance, i.e. u_S ≡ 0, simplifying Eq. (1), while the resistant strain carries an evolving trait u_R ≔ u following Darwinian dynamics along the fitness gradient, defined as g(u; C_b, N_S, N_R)= $\frac{\partial f_{R} (C_{b}, N_{S}, N_{R}, u)}{\partial u} .$

In Eqs. (1) and (2), r_S and r_R denote the intrinsic growth rates of susceptible and resistant strains, respectively, and d represents their natural death rate. The carrying capacity K defines the maximal population size sustainable by available resources. Parameters a_SR and a_RS quantify inter-strain competition, while intra-strain competition is captured by the self-interaction coefficients a_RR and a_SS, both fixed to 1. $E_{\max}$ denotes the maximal antibiotic killing rate. The half-maximal effective concentrations $E C 5 0_{S_{0}}$ and $E C 5 0_{R_{0}}$ correspond to the baseline drug levels producing 50% of $E_{\max}$ for susceptible and resistant bacteria, respectively⁴¹. The serum amikacin concentration C_b, computed by the PBPK module, defines the selective environment shaping resistance evolution, linking PK to bacterial adaptation. The resistant strain incurs a metabolic cost c associated with expressing the resistance trait u, leading to a reduction in growth rate, in line with evidence showing an exponential decline in bacterial growth with increasing resistance level⁴². The corresponding increase in EC50 with resistance level is also experimentally supported⁴³, and in our model is modulated by the parameter k_u, which determines how strongly the effective EC50 increases as the resistance trait rises.

The complete model is given below, with σ denoting the evolutionary rate.

\{\begin{matrix} {\dot{N}}_{S} = f_{S} (C_{b}, N_{S}, N_{R}, 0) N_{S} \\ {\dot{N}}_{R} = f_{R} (C_{b}, N_{S}, N_{R}, u) N_{R} \\ \dot{u} = σ \frac{\partial f_{R} (C_{b}, N_{S}, N_{R}, u)}{\partial u} = σ [- r_{R} c e^{- c u} (1 - \frac{N_{R} + a_{R S} N_{S}}{K}) + \frac{E_{\max} C_{b} E C 50_{R_{0}} k_{u} e^{k_{u} u}}{{(E C 50_{R_{0}} e^{k_{u} u} + C_{b})}^{2}}] \end{matrix})

Table 1 summarizes the PD and eco-evolutionary model parameters, their biological interpretation, units, and the calibrated ranges used in the present study.

Table 1.

PD parameters: definitions, calibrated ranges, and units

Symbol	Description	Range	Unit
$E C 5 0_{S_{0}}$	Baseline C_b producing 50% of $E_{\max}$ on N_S	[6, 40]	[mg/L]
$E C 5 0_{R_{0}}$	Baseline C_b producing 50% of $E_{\max}$ on N_R	[6, 40]	[mg/L]
$E_{\max}$	Maximum drug-induced effect achievable	[1.0, 10.0]	[1/day]
c	Metabolic cost in adaptation	[0.02, 0.6]	–
r_S	Growth rate of N_S	[0.4, 1.5]	[1/day]
r_R	Growth rate of N_R	[0.1, 0.7]	[1/day]
k_u	Pharmacodynamic sensitivity	[0.2, 1.0]	–
σ	Evolutionary rate	[0.05, 0.3]	[1/day]
d	Death rate of N_S and N_R	[0.05, 0.18]	[1/day]

Open in a new tab

Mechanistically informed calibration bridges neonatal physiology and eco-evolutionary dynamics

Since model parameters, especially in a newly proposed model, critically influence predictive accuracy and must capture inter-patient variability, we performed a dedicated parameter identification procedure. Specifically, PD parameters in Table 1 were identified through a constrained, behavior-driven calibration enforcing biologically expected bacterial dynamics under clinically observed drug exposure, as detailed in the Methods and Supplementary Note 2.

Across GA–PNA strata, calibrated baseline potencies $E C 5 0_{S_{0}}$ and $E C 5 0_{R_{0}}$ clustered near the identity line ( $E C 5 0_{S_{0}} \approx E C 5 0_{R_{0}}$ ; Supplementary Fig. S1), indicating that no artificial baseline resistance was introduced. During antibiotic exposure, the effective potency of the resistant strain evolves as $E C 5 0_{R} = E C 5 0_{R_{0}} e^{k_{u} u},$ so that increases in the adaptive trait u reduce apparent drug efficacy in the PD killing term $\frac{E_{\max} C_{b}}{E C 5 0_{R} + C_{b}}$ , effectively raising the MIC. This resistance gain is offset by the metabolic cost of adaptation, represented by the growth penalty r_Re^−cu, which slows down replication in drug-free conditions. Accordingly, $E C 5 0_{S_{0}}$ and $E C 5 0_{R_{0}}$ serve as interpretable MIC surrogates summarizing the calibrated potency landscape across the cohort. The remaining parameters (k_u, σ, c, r_S, r_R, $E_{\max}$ ) govern temporal evolution of growth and adaptation. Analytical insights into the interplay between k_u and σ, which jointly determine the strength and speed of adaptation, are provided in Supplementary Note 1. This coupling was accounted for during calibration by applying weak priors and bounded ranges to ensure realistic rise-decay dynamics of u.

Consistently with developmental PK trends^1,8, the overall selection pressure for resistance decreased with increasing maturity (Fig. 2, and Supplementary Fig. S2).

In more mature neonates, faster amikacin clearance shortens exposure and deepens troughs, thus reducing the time beneficial to resistance development. Accordingly, the selection-onset probability, defined as the fraction of neonates with a positive early-time selection gradient g(u; C_b, N_S, N_R) > 0, (i.e., conditions favoring an initial rise in the adaptive trait u), declined systematically with both GA and PNA. The resulting eco-evolutionary phase maps $(u, \log_{10} (C_{b} / E C 5 0_{R_{0}}))$ reveal a clear maturational pattern. Preterm neonates, owing to immature renal clearance, maintain elevated drug concentrations for longer periods, which in turn broadens the regions of positive selection (with probability P(g > 0) ≈ 1). Conversely, with increasing GA and PNA, these regions progressively narrow, reflecting the maturation of glomerular filtration and the resulting increase in antibiotic elimination. Thus, preterm infants face extended selective windows favoring resistance, whereas term neonates have briefer, less selective exposures that facilitate the recovery of susceptible populations^44–46.

To illustrate how antibiotic exposure reshapes the eco-evolutionary fitness landscape, we evaluated the selection gradient g(u; C_b, N_S, N_R) along the PBPK trajectory of a representative preterm neonate. Three time points spanning peak, mid-, and late-interval exposures reveal a transition from drug-dominated (g > 0) to near-neutral or negative selection as C_b declines. Figure 2 shows the corresponding fitness surfaces over $(u, \log_{10} (C_{b} / E C 5 0_{R_{0}}))$ , where the normalized exposure axis centers the scale at $C_{b} = E C 5 0_{R_{0}}$ to symmetrize sub- and supra-inhibitory ranges. Immediately after dosing, high C_b produces broad regions of positive selection favoring resistance, which progressively shrink and invert as exposure decreases. This cyclic alternation between drug-driven selection and ecological recovery reproduces the expected adaptive dynamics under intermittent antibiotic treatment, confirming the model’s mechanistic validity.

Data-driven modeling captures neonatal serum creatinine trajectories across heterogeneous clinical cohorts

We used data from three complementary real-world neonatal cohorts: AMICREA-CI, AMICREA-TH, and CREA-AKITH including longitudinal sCr measurements from 706, 56, and 869 neonates, respectively, to develop predictive models of sCr dynamics from which eGFR, a key renal function parameter in the PBPK model, was derived. These cohorts were characterized by inotrope/ibuprofen co-administration (AMICREA-CI), therapeutic hypothermia (AMICREA-TH), and AKI with therapeutic hypothermia (CREA-AKITH) and differ in terms of neonate’s maturity range, the presence of comorbidities, and concomitant interventions including other pharmacological treatments, thus providing a diverse basis for model development and evaluation (Table 7).

Table 7.

Clinical characteristics of patients in the three neonatal cohorts

Clinical variables median [range]	AMICREA-CI	AMICREA-TH		CREA-AKITH
	TDM = 3573, n = 709	TDM = 158, n = 56	p-value	sCr = 5444, n = 869	p-value
Gestational age (weeks)	34 [24–42]	38 [35–41]	<0.001	40 [34–43]	<0.001
Postnatal age (days)	2 [1–30]	1.5 [1–11]	<0.001	3 [1–10]	0.014
Current weight (grams)	2102 [385–4650]	3038 [1910–4770]	0.003	3335 [1750–6230]	<0.001
Serum creatinine (mg/dL)	0.87 [0.40–2.87]	0.86 [0.28–2.21]	<0.001	0.76 [0.08–4.10]	<0.001
Blood culture test
Positive	TDM = 498, n = 101	–		–
Negative	TDM = 3075, n = 608	–		–
Comorbidities
Perinatal Asphyxia	–	TDM = 158, n = 56		–
Acute Kidney Injury	–	–		TDM = 908, n = 127
Co-interventions
Therapeutic Hypothermia	–	TDM = 144, n = 55		TDM = 5444, n = 869
Respiratory support	TDM = 2207, n = 442	TDM = 148, n = 54		–
Ventilation	TDM = 1516, n = 310	TDM = 113, n = 43		–
Inotropes	TDM = 603, n = 134	TDM = 65, n = 31		–
Ibuprofen	TDM = 173, n = 47	–		–

Open in a new tab

The comparisons of continuous variables GA, PNA, BW, and sCr measurements between three datasets are performed using two-sided Mann–Whitney U test. ‘—’ indicates no available TDM measurements.

n neonates per group.

In Fig. 3a–c, sCr data from the three cohorts are shown with the corresponding empirical centile curves (p10–p90) over PNA. The distinct postnatal sCr trajectories across cohorts reflect their specific pathophysiology and clinical conditions. As shown in Fig. 3a, in the AMICREA-CI cohort involving newborns with GA spanning 24–42 weeks, the median peak value (≈0.8–0.9 mg/dL) occurs 2.3 days after birth, followed by a gradual decrease, approaching ≈0.4 mg/dL at the end of the observed postnatal period (approximately after 24.7 days). This phenomenon reflects established neonatal renal physiology in the first days of life, driven by maternally transferred creatinine at birth^47–49 and immature tubular handling^18,26, with creatinine subsequently cleared as renal maturation progresses. The AMICREA-TH and CREA-AKITH cohorts involved a much smaller range of GA (34–42 weeks) and PNA up to 10 days, because TH is only established for use in (near) term newborns during the first three days after birth. A systematic review proved this time window as sufficient to cover the TH-specific sCr pattern⁵⁰.

In the AMICREA-TH cohort (Fig. 3b), the peak appears less pronounced reflecting the combined effect of reperfusion injury attributed to perinatal asphyxia and TH. In the CREA-AKITH cohort (Fig. 3c), a marked increase in sCr levels is observed, reflecting prolonged kidney injury. As highly vaso-reactive organs, the kidneys are particularly susceptible to oxygen deprivation, and hypoxic-ischemic damage frequently results in AKI as part of the perinatal asphyxia syndrome⁵¹. The cohort exhibits high between-subject variability, with median sCr at PNA day 1 of about 0.92 mg/dL, which declines to 0.57 mg/dL by day 3 and continues to decrease at a much slower rate to reach the value of 0.39 mg/dL on day 10.

Three state-conditional long short-term memory (LSTM) based deep neural networks were featured to learn longitudinal dependencies in the above-mentioned pathophysiological time-sequence data to infer sCr dynamics. Following the architecture proposed by Hochreiter and Schmidhuber⁵², a recurrent neural network with a gated LSTM structure was implemented. The network consists of three consecutive LSTM layers to mitigate vanishing and exploding gradient issues during multistage backpropagation, followed by a dropout layer and a final fully connected layer designed for regression. Implementation details are provided in the Methods. In each workflow, a three-fold cross-validation procedure, stratified by GA and PNA at the patient level, was employed to train and test each model.

Comparative validation reveals superior fit to observed postnatal trajectories

We validated our data-driven LSTM models against the mechanistic population model proposed by Krzyzanski et al.¹⁸, which describes postnatal sCr dynamics in neonates as a function of PNA and GA. To account for potential bias arising from heterogeneity among the study cohorts, the same three-fold cross-validation partitions were used to train from scratch and to validate the Krzyzanski et al. model and our LSTM-based neural network. This approach allowed full parameter recalibration and mitigated possible bias inherent in the original datasets. The model in ref. ¹⁸ explicitly describes the transient back-flow of creatinine from the renal tubules to the plasma during the first few days after birth. However, in neonates with AKI or undergoing PATH, these peaks are smoothed (see Fig. 3b, c), resulting in increased residuals, while our models provide a more unbiased prediction of sCr trajectories in this early postnatal window. Figure 3d–f we report, for each PNA day in the dataset, the distribution of residuals obtained by aggregating the predictions across the three validation folds. We evaluated predictive performance using the mean prediction error (MPE) and the mean absolute prediction error (MAPE), formally defined in the Methods. Our models show a slight underestimation bias, which is nonetheless smaller in both AMICREA-TH and CREA-AKITH cohorts, with an MPE of −0.044 mg/dL vs 0.088 mg/dL and −0.059 mg/dL vs 0.083 mg/dL with respect to the model of Krzyzansky et al. In all three cohorts, our model registered increased precision. In the AMICREA-CI cohort, a MAPE of 0.154 mg/dL vs 0.168 mg/dL was achieved (Fig. 3d), corresponding to a mean reduction of 8.3%, with reduced residuals across the entire PNA range of 1–30 days. During and after TH, our model registers a MAPE of 0.176 mg/dL vs 0.271 mg/dL in the AMICREA-TH, corresponding to a mean MAPE reduction of 35.1%, and 0.223 mg/dL vs 0.398 mg/dL in CREA-AKITH cohorts, corresponding to a mean reduction of 44.0%, respectively (Fig. 3e, f). These results confirm the improvement in both accuracy and precision in sCr prediction across heterogeneous clinico-pathological neonatal conditions.

The digital-twin framework accurately recapitulates observed amikacin PK and blood culture positivity patterns

To provide an integrated assessment of the model’s predictive performance, we jointly evaluated the ability of the DT framework to reproduce both the observed amikacin PK and the clinical patterns of blood culture positivity in the neonatal cohorts. This was achieved by analyzing the variation of simulated bacteria colony-forming units (CFUs) over the course of therapy.

Recapitulation of observed amikacin pharmacokinetics

The fully calibrated PBPK-PD model was implemented in MATLAB^® SimBiology and incorporated time-varying parameters, paired with the real patient characteristics captured across the three cohorts. Renal clearance was scaled linearly with body weight and eGFR, while organ flows were distributed proportionally to cardiac output (CO).

The simulated time horizon was aligned with the actual amikacin administration schedule and extended up to 24 h after the last TDM measurement or infusion recorded in the NICU. Consequently, the total simulated therapy duration varied substantially across patients, from ~50 to 600 h in AMICREA-CI cohort. As illustrated in Fig. 4a, c, for some randomly selected patients, the PBPK model accurately reproduced amikacin concentrations throughout the treatment period, showing no evidence of temporal drift. This confirms the model’s capability to capture the relevant pathophysiological processes governing amikacin disposition, even in neonates with pathological conditions and co-medications.

Fig. 4 — Amikacin dynamics simulated for randomly selected patients during the NICU stay by the PBPK model, with a comparison between predicted concentrations (green asterisks) and measured values (red dots) at the clinical sampling times recorded in the AMICREA-CI (a) and AMICREA-TH (c) real-world cohorts. Distribution of the Average Fold Error, stratified by GA and concomitant inotrope/ibuprofen administration in the AMICREA-CI cohort (b), and by GA and concomitant TH in the AMICREA-TH cohort (d). e Evaluation of simulated CFU variation distribution between observed blood culture proven infection outcomes in the real-world cohort AMICREA-CI.

By updating clearance and organ blood flows according to longitudinal eGFR, CO, GA, and PNA, the PBPK simulations reproduced the observed concentration-time profiles with high fidelity, particularly during the first weeks of life when renal and circulatory maturation are most pronounced. The adaptive parameterization allowed smooth transitions between physiological states without discontinuities at dosing or sampling times. After the full simulation, predicted amikacin concentrations are interpolated at the exact time points of measured concentrations.

To enable a clinically meaningful interpretation of model performance, the distribution of the Average Fold Error (AFE) across patient subgroups and cohorts is shown in Fig. 4b, d, while observed amikacin plasma concentration ranges, stratified by GA and relevant clinical modifiers, are reported in Table 2 to contextualize the reported prediction errors. Across the AMICREA-CI cohort (Table 2a), observed concentrations span wide ranges, particularly in very preterm neonates (<32 wk) with aggregated IQR (54.6–88.7) mg/L. In this context, Mean Absolute Error (MAE) values in the range of 3.6–7.6 mg/L correspond to relative deviations that are modest when interpreted against the observed exposure variability, supporting the clinical acceptability of the model predictions. Consistently, AFE values are centered around unity across most strata (Table 2), indicating an overall absence of systematic bias, with mild under- or over-prediction emerging in specific subgroups characterized by co-administrations. Notably, co-administration of inotropes and/or ibuprofen is associated with both altered exposure ranges and changes in predictive error, reflecting the additional physiological complexity introduced by these clinical conditions. At the cohort level, a median AFE of 0.95 [95% CI: 0.51, 1.91], and a median MAE of 7.72 mg/L [95% CI: 0.88, 15.25] were achieved across 709 patients and 3575 TDM measurements. Both metrics are formally defined in the Methods.

Table 2.

Measured amikacin concentration ranges, reported as the difference between maximum and minimum observed values with relative IQR (mg/L), together with prediction error metrics AFE and MAE (mg/L)

(a) AMICREA-CI cohort
GA (wk)	Co-administrations	Amikacin range (mg/L)	AFE	MAE (mg/L)
<32	None	71.6 (53.6–89.6)	1.05	7.57
	Ino only	61.1 (48.9–73.4)	0.91	4.49
	Ibu only	42.9 (27.1–58.8)	1.14	5.44
	Both	70.0 (55.2–84.8)	0.91	3.57
32–37	None	60.1 (43.5–76.7)	0.98	8.62
	Ino only	46.4 (31.9–60.9)	0.87	6.69
	Ibu only	39.0 (23.4–54.6)	1.02	8.08
	Both	37.0 (18.5–55.5)	1.32	8.32
>37	None	49.1 (33.9–64.3)	0.92	7.62
	Ino only	53.2 (41.7–64.7)	0.98	5.36
	Ibu only	37.5 (21.3–53.7)	1.02	7.59

(b) AMICREA-TH cohort
GA (wk)	TH	Amikacin range (mg/L)	AFE	MAE (mg/L)
32–37	No	9.7 (6.1–13.3)	0.46	5.86
	Yes	18.9 (15.2–22.6)	0.43	5.31
> 37	No	26.5 (18.2–34.8)	0.51	5.54
	Yes	37.0 (32.8–41.2)	0.66	3.96

Open in a new tab

a: Stratification by GA and concomitant inotrope and/or ibuprofen administration in the AMICREA-CI cohort. b: Stratification by GA and TH in the AMICREA-TH cohort.

In the AMICREA-TH cohort (Table 2b), observed amikacin concentration ranges were narrower, and associated with consistently lower MAE values (3.9–5.9 mg/L). AFE values below 1 indicate conservative yet sufficiently accurate predictions for most TH-treated patients (GA >37), with an AFE of 0.66 and an MAE of 3.9 mg/L. At the cohort level, the model showed a median AFE of 0.58 [50% CI: 0.41, 0.71] and a median MAE of 4.17 mg/L [50% CI: 2.49, 6.54] across 56 patients and 158 TDM measurements.

The relatively narrow confidence intervals and the lower AFE, indicating a tendency toward underprediction, reflects the limited data support of this cohort and the restricted early postnatal window, which coincides with pronounced transient renal maturation. As described in the Methods, the risk of overfitting in this setting was mitigated by the use of a reduced LSTM architecture and a simplified input feature set for renal function modeling, which improves robustness but limits the representation of short-term, patient-specific variability in renal clearance that directly scales amikacin elimination in the PBPK model. Consequently, unmodeled sources of renal-function variability in this clinical setting, such as hemodynamic instability and supportive intensive care interventions, may contribute to more sustained observed concentrations than predicted.

Recapitulation of blood culture positivity patterns

Longitudinal blood-culture results collected during therapy and available for the AMICREA-CI cohort were used for a posteriori validation of the PD component, enabling assessment of the model’s ability to distinguish between positive and negative infection states along individual patient trajectories.

Figure 4e shows that the model-predicted relative changes in CFU ( $Δ \log_{10}$ CFU), based on the real-world antibiotic dosing implemented in the AMICREA-CI cohort, are clearly stratified according to blood culture-proven infection outcomes (p < 10⁻¹⁶ by two-tailed Wilcoxon rank-sum test). Notably, this stratification emerges despite the model not being calibrated on patient-specific bacterial counts, but instead initialized using a standardized, empirically derived inoculum based on established neonatal sepsis models. Patients with confirmed bloodstream infection exhibit higher predicted CFU trajectories than culture-negative cases, indicating that the PD module captures biologically meaningful differences in bacterial load dynamics and reflects underlying infection status, thereby supporting its potential clinical relevance.

The virtual cohort reproduces maturational and pathophysiological neonatal patterns

We built a virtual neonatal cohort (N = 1000) spanning GA 24–42 wk and PNA 1–28 d (see Supplementary Note 3), to reflect and extend the renal and cardiovascular variability observed in the real neonatal cohorts. Each subject combined maturational attributes (GA, PNA, BW, and body height–Ht) with clinical modifiers including TH, ibuprofen (Ibu) exposure, inotrope (Ino) support, and PDA. Predicted physiological variables such as eGFR, sCr, heart rate (HR), stroke volume (SV), and CO, were parameterized to reflect the heterogeneity observed in NICU populations. Cohort-level maturational characteristics (GA, PNA, BW, Ht) are summarized in Supplementary Fig. S3, while predicted CO illustrates the resulting physiological diversity across individuals.

Realism of renal function across pathophysiological conditions

We assessed the impact of each clinical condition and co-medication on the distribution of eGFR, and on its precursor sCr, as predicted by the LSTM-based deep neural network. Following the prematurity ranges defined by Glass et al.⁵³, Fig. 5 shows GA-stratified boxplots across three groups, Extremely/Very Preterm (GA < 32 wk), Moderate/Late Preterm (32–37 wk), and Term (>37 wk), and three PNA ranges (<5 d, 5–15 d, >15 d). Consistent with the functional maturation of glomeruli and renal tubules, sCr decreases and eGFR increases with advancing GA and PNA. This maturational rise in eGFR persists in neonates with AKI but at lower absolute values, consistent with impaired renal function⁵⁴. TH-treated neonates show attenuated maturational trends due to the narrow exposure window and the absence of very preterm subjects, yet exhibit the expected sCr elevation linked to TH⁵⁵. Inotrope and ibuprofen treatments display heterogeneous but generally higher sCr, suggesting illness severity and mild nephrotoxic contributions⁵⁶. Similarly, PDA elevates sCr by altering systemic and renal hemodynamics through a redistribution of CO, in line with reduced eGFR and impaired tubular handling⁵⁷. Median eGFR and sCr across AKI and TH stratifications are summarized in Table 3a, b, respectively. Overall, the model reproduces physiologically plausible maturational and pathological trends consistent with neonatal renal physiology.

Fig. 5 — Effects of non-maturational factors including AKI (a), TH (b), PDA (c), and co-administrations of inotropes/ibuprofen (d). Effects of maturational factors (PNA) on sCr (e, f) and eGFR (g) values. Parts of this figure were created with BioRender.com.

Table 3.

Median renal function indicators by GA and clinical modifiers

a) eGFR (mL/min/1.73 m²) by AKI status			b) sCr (mg/dL) by TH status
GA (wk)	AKI–No	AKI–Yes	GA (wk)	TH–No	TH–Yes
<32	0.928 (n = 263)	0.701 (n = 137)	<32	0.735 (n = 400)	–
32–37	0.962 (n = 267)	0.554 (n = 63)	32–37	0.738 (n = 306)	0.934 (n = 24)
≥37	1.166 (n = 253)	0.617 (n = 17)	≥37	0.718 (n = 175)	0.844 (n = 95)

Open in a new tab

"—” indicates strata not represented by the clinical modifier (i.e., TH not applied in very preterm infants).

n neonates per group.

Realism of heart rate dynamics across maturational and clinical conditions

In the virtual cohort, HR is generated through GA- and PNA-dependent baselines, with explicit modifiers for TH and inotrope support. For non-cooled neonates, the mean HR, denoted μ_HR, follows a linear GA trend, consistent with findings reported by Hsu et al.⁵⁸: μ_HR = 202.7 − 1.92 ⋅ GA, which yields ~149 bpm at 28 wk and ~126 bpm at term. Individual HR values are drawn from a normal distribution centered on μ_HR with physiologic bounds. To model the bradycardic response to TH, we drew on the findings of Elstad et al.⁵⁹, which reported a characteristic reduction in HR over the first 72 h of TH in term infants. Accordingly, for cooled term neonates (GA ≥36 wk) we implemented a time course as in ref. ⁵⁹ by applying a time-varying offset that depresses the baseline toward target medians of ~92 bpm at ~12 h and ~97 bpm at ~72 h. Inotrope exposure produces additive increases of +15 bpm under TH (Fig. 6c) and +5 bpm otherwise. This construction preserves maturational trends outside TH while reproducing the expected bradycardic response during TH.

Fig. 6 — a Modeled PDA prevalence as a function of PNA across GA strata, with and without Ibu exposure. Solid lines indicate baseline PDA probability, and dashed lines show the Ibu-induced reduction. The insert illustrates the modeled spontaneous attenuation of PDA severity from day 1 to day 5. b Heart rate (HR) trajectories in cooled term neonates (GA >36 wk) over 0–96 h. Median (solid line) and interquartile range (IQR, shaded area) replicate the characteristic bradycardic response during TH and recovery during rewarming. c Effect of inotropic support on HR during TH (12–24 h), showing the expected increase in HR under Ino exposure. Parts of this figure were created with BioRender.com.

When stratifying by GA and PNA, the simulated HR closely matches the expected clinical patterns. Indeed, among term neonates (GA ≥ 37 wk) within 72 h, cooled infants had a median HR of 95 bpm [89–102] (n = 106), versus 132 bpm [125–140] (n = 8) in non-cooled peers. Windowed analyses centered at ~12 h and ~72 h in cooled term neonates reproduce the medians in ref. ⁵⁹ (92 and 97 bpm) with overlapping interquartile range (IQRs) and reported 95% confidence interval, as shown in Fig. 6b. For completeness, cohort-level medians by TH status were 94 bpm [88–102] in TH (n = 119) and 138 bpm [127–148] in non-TH (n = 881). The higher non-TH median reflects cohort composition (a predominance of preterm infants, who have higher baseline HR), rather than a mismatch of the maturational trend. Accordingly, HR in the preterm non-TH subgroup was 141 bpm [132–151] (n = 687), consistent with the GA-dependent baseline.

Realism of cardiovascular indicators and transitional hemodynamics

To ensure physiologic plausibility of CO in our virtual neonatal population, we explicitly modeled HR and SV as functions of GA, PNA, and clinical modifiers.

Stroke volume index (SVI, mL/m²) was assigned from GA- and PNA-dependent baselines (SVI_28wk = 22 mL/m², slope: −0.1 mL/m²/wk), bounded between 15 and 30 mL/m². Early PNA (≤2 d) conferred a +1 mL/m² increment, whereas TH in term neonates reduced SVI by 1 mL/m²; inotrope use increased SVI by 1 mL/m². PDA was modeled as highly prevalent at very low GA, with an initial probability p₀ that decreased with increasing GA (p₀ = 0.9 at GA < 28 wk down to p₀ = 0.01 at term). The probability of PDA persistence declined exponentially with PNA according to p = p₀e^−PNA/9, reflecting spontaneous closure over the first week of life. Ibuprofen exposure reduced incidence by 40% (see Fig. 6a).

A PDA event was then sampled as a Bernoulli variable with probability p. For neonates in whom PDA occurred, severity at presentation was assigned probabilistically: 50% mild, 35% moderate, and 15% severe. Severity was subsequently attenuated after day 3 to reflect the natural progressive tightening of the duct. When administered, ibuprofen further reduced severity and the duct remained patent. The resulting PDA severity (none/mild/moderate/severe) was mapped to a multiplicative increase in SVI of +5%, +10%, or +18%, respectively.

Body surface area (BSA, m²) was calculated by the Haycock formula:⁶⁰

B S A = 0.024265 \cdot {B W}^{0.5378} \cdot {H t}^{0.3964}

and SV (mL) as SV = SVI ⋅ BSA.

CO (L/min) was then obtained as $C O = \frac{H R \cdot S V}{1000}$ . Finally, cardiac index was calibrated to published reference mean values (2.3–2.7 L/min/m² across GA strata)⁵⁸, with GA-bin-specific scaling during the 72–96 h window (non-TH neonates). This combination yielded realistic transitional physiology, including higher CO in PDA-positive preterms and depressed CI under TH, consistent with clinical expectations.

To validate hemodynamic realism in our non-cooled neonates, we compared our values with the GA-specific reference means of CO, CI, HR, and SV reported in ref. ⁵⁸ at 72–96 h. In our model, the overall CI median was 2.60 L/min/m² (n = 92), matching the target range of 2.55–2.60 L/min/m². When stratified by GA, our simulated mean± standard deviation values closely replicated the reference values for all variables (Table 4), confirming preservation of the correct maturational trajectory of CO, CI, HR, and SV.

Table 4.

Comparison of simulated and reference hemodynamic values at 72–96 h in non-cooled neonates⁵⁸

GA (wk)	Source	CO (L/min)	CI (L/min/m²)	HR (bpm)	SV (mL)
≤28	Our study	0.24± 0.02	2.62± 0.26	148± 8.5	1.64± 0.14
	Reference	0.23± 0.03	2.31± 0.26	149± 11.4	1.56± 0.28
29–30	Our study	0.25± 0.01	2.51± 0.12	145± 15.4	1.73± 0.16
	Reference	0.29± 0.06	2.45± 0.24	145± 8.7	1.99± 0.44
31–32	Our study	0.33± 0.03	2.69± 0.25	142± 10.7	2.35± 0.19
	Reference	0.35± 0.07	2.69± 0.36	142± 8.9	2.53± 0.47
33–34	Our study	0.35± 0.07	2.54± 0.31	142± 7.5	2.48± 0.50
	Reference	0.35± 0.07	2.54± 0.32	142± 15.7	2.49± 0.60
35–36	Our study	0.40± 0.05	2.64± 0.32	136± 10.6	2.96± 0.24
	Reference	0.43± 0.08	2.64± 0.33	136± 13.5	3.22± 0.70
37–38	Our study	0.46± 0.11	2.49± 0.19	131± 10.0	3.50± 0.88
	Reference	0.47± 0.10	2.49± 0.39	131± 11.9	3.67± 0.72
39–41	Our study	0.61± 0.07	2.60± 0.30	126± 12.8	4.87± 0.43
	Reference	0.53± 0.14	2.60± 0.54	126± 12.1	4.24± 0.89

Open in a new tab

Reported values are mean± standard deviation.

Evolutionary digital-twin optimization enables personalized and effective neonatal dosing

Across the neonatal virtual population, the optimization algorithm identified safe and effective individualized amikacin regimens (less than 7 days on average) terminating once a sustained bacterial reduction persisted for at least 3 consecutive days.

Initial bacterial loads were defined to avoid assuming a priori a marked predominance of either susceptible or resistant subpopulations and were chosen to be consistent with the inoculum used during model calibration. Therefore, the two populations were initialized at comparable levels, with $N_{S_{0}} = 8.8 \times 1 0^{5}$ and $N_{R_{0}} = 2.2 \times 1 0^{5}$ . Importantly, since optimization and stopping criteria are driven by relative bacterial reduction dynamics rather than absolute CFU values, the resulting dosing strategies are robust to uncertainty in the initial inoculum.

The optimizer dynamically refines therapy according to the evolving physiological and microbiological states of each virtual neonate, effectively serving as a personalized DT.

Iterative single-cycle optimization ensures robust and adaptive PK control under safety constraints

At each cycle, the optimizer solves a constrained nonlinear problem over the feasible dose range (10–25 mg/kg, 1 h infusion), targeting the PK-PD ratio of $C_{\max} / M I C \geq 8$ (standard target for efficacy, with $C_{\max}$ denoting the peak serum amikacin concentration²⁴), and enforcing an interdose interval $T_{C_{\min}} \leq 48$ h, where $T_{C_{\min}}$ denotes the first post-infusion time at which amikacin serum concentration falls below a fixed safety threshold of 3 mg/L. As nephrotoxicity is not explicitly modeled in the present framework, toxicity is addressed through well-established exposure-based safety thresholds; accordingly, regimens with $C_{\max} \geq 35$ mg/L were penalized as overexposure²⁴, and any exceeding 60 mg/L were rejected (hard cap), in line with current clinical practice for aminoglycoside dosing in neonates. The terminal PK state of each cycle defines the initial conditions of the next, ensuring temporal continuity and enabling adaptive control as patient physiology evolves. At each $T_{C_{\min}}$ the optimizer re-evaluates the patient state and adjusts the next dosing plan.

Between cycles, the virtual neonate’s hemodynamic state and physiological parameters are updated according to the age- and condition-dependent PBPK model, while the bacterial susceptibility profile adapts according to the resistance-trait dynamics (u in Eq. (3)) in the PD module. Hence, the effective MIC for the next cycle is computed via a Hill-type saturation function:

M I C (t + Δ t) = M I C (t) + ({M I C}_{\max} - M I C (t)) \frac{u {(t)}^{h}}{u_{50}^{h} + u {(t)}^{h}}

where MIC(0) = MIC₀ is the baseline susceptibility (see Supplementary Note 3) and ${M I C}_{\max} = 8$ mg/L represents the upper limit of phenotypic resistance reachable in the neonatal population under selective pressure. The trait u is dimensionless and identifiable only up to a scale factor, as it appears in exponential and multiplicative terms within the eco-evolutionary dynamics. Accordingly, the midpoint u₅₀ = 300 is chosen as a numerically convenient scaling constant that, together with the PD parameters σ, k_u and their composite sensitivities σk_u and $σ k_{u}^{2}$ (Supplementary Note 1) calibrated at the population level, yields slow MIC drifts over the short (about 7 days) treatment horizon. This choice fixes the internal scale of u without altering model behavior. Equivalently, the formulation can be expressed using a normalized trait. However, a rigorous normalization of the resistance trait and its implications for model scaling and parameter identifiability would require a dedicated structural analysis, which is beyond the scope of the current study. This formulation yields a smooth, biologically consistent rise in MIC as u increases, reflecting the nonlinear link between resistant-subpopulation enrichment and effective population-level susceptibility⁶¹. Together, these mechanisms close the adaptive feedback loop, re-optimizing dosing after each cycle to match the evolving patient physiology and resistance state, ensuring sustained PK-PD target attainment under gradual eco-evolutionary change.

Optimization achieves high PK/PD target attainment while maintaining safety limits

To evaluate global performance beyond the optimization criteria, additional PK/PD indices such as AUC/MIC and exposure-related $C_{\max}$ thresholds were computed a posteriori. Across all MIC strata (2–3, 3–4, 4–6, and 6–8) mg/L, the evolutionary DT optimizer achieved high or complete attainment of PK/PD benchmarks (Fig. 7):

$T_{C_{\min}} \leq 48$ h was satisfied in 100% of cases across all strata.
AUC/MIC ≥ 21.4 (for bacteriostasis) was reached in all patients, and AUC/MIC ≥ 62.5 (for 1- $\log_{10}$ CFU/mL reduction) in 94–98% for MIC ≤ 6 mg/L, decreasing to 58% for MIC 6–8 mg/L.
$C_{\max} \geq 15$ mg/L was reached in all patients, $C_{\max} \geq 24$ mg/L was achieved in 72% (MIC 2–3 mg/L) and 100% otherwise, while the soft cap ( $C_{\max} \geq 35$ mg/L) was reached selectively for higher MICs (99–100%).
$C_{\max} / M I C \geq 8$ was achieved in all neonates with MIC ≤ 6 mg/L; for MIC 6–8 mg/L, 95% remained within $6 \leq C_{\max} / M I C \leq 8$ .

The apparent underperformance of the $C_{\max} / M I C \geq 8$ target for isolates with MIC values of 6–8 mg/L does not reflect a flaw in the optimizer, but rather the imposed pharmacological constraints. With a soft cap at $C_{\max} = 35 m g / L$ , the highest achievable efficacy ratio is 35/MIC, meaning that the target $C_{\max} / M I C = 8$ can only be attained for MIC ≤ 4.4 mg/L. For higher MIC values (6–8 mg/L), the optimizer is structurally unable to meet the target without exceeding safe concentrations. Even under the strict upper bound of $C_{\max} = 60 m g / L$ , the ratio would reach at most 7.5 (for MIC = 8) or require 48 mg/L for MIC = 6, approaching toxic exposure. Accordingly, the optimizer prioritizes safety, maintaining $C_{\max} / M I C$ ratios between 6 and 8 in about 95% of cycles rather than forcing unsafe peaks. This analysis confirmed that the optimized regimens maintained acceptable therapeutic performance across the entire susceptibility spectrum. Beyond overall performance, representative dose and dosing-interval recommendations within clinically relevant virtual patient subgroups (Supplementary Note 4) also reveal consistent and clinically interpretable patterns, showing that optimized regimens remain within established neonatal dosing practices while allowing adaptive adjustments over the course of therapy in response to evolving PBPK-PD dynamics.

To quantify the dose–schedule trade-off visible in Fig. 7c, we fitted a linear mixed-effects model with a patient-specific random intercept:

T_{C_{\min}} ~ {Dose}_{opt} \times MIC + GA + BW + eGFR + (1 ∣ PtID) .

The fixed–effects estimates show the expected positive association between dose and interdose interval:

β_{Dose} = 0.447 h per mg / kg (95 % C I 0.377 - 0.517, p < 1 0^{- 34}),

implying that moving from 10 to 25 mg/kg is associated with an increase of ≈0.447 × 15 ≈ 6.7 h in $T_{C_{\min}}$ (on average).

Neither the main effect of MIC nor the Dose × MIC interaction was significant (β_MIC = − 0.047 h, p = 0.90; β_Dose×MIC = − 0.0125 h, p = 0.42), indicating that $T_{C_{\min}}$ is essentially constant across the susceptibility range 2–8 mg/L. That is, higher MIC values shift the required dose (Fig. 7f), but leave the fundamental dose-interval trade-off unchanged, as reflected by the nearly parallel MIC-specific regression lines in Fig. 7c. This finding is complemented by the aggregated analysis reported in Supplementary Note 4, which shows that dosing intervals are predominantly modulated by maturational status, renal impairment, and disease-related conditions. Moreover, the physiological covariates behave as expected: (i) GA is associated with shorter intervals (β_GA = − 0.218 h per week, 95% CI −0.293 to −0.142), consistent with faster clearance in more mature neonates; (ii) BW shows a small positive effect (β_BW = 3.88 × 10⁻⁴ h per gram; ≈0.39 h per additional kilogram) and (iii) eGFR is strongly negative (β_eGFR = − 11.64 h on the model’s scale, p ≈ 0), reflecting shorter intervals with better renal function. The random intercept standard deviation is 2.47 h (inter-patient heterogeneity), and the residual standard deviation is 2.02 h (inter-occasion variability).

Pharmacodynamic outcomes reveal emergent sensitive dominance and ecological patterns

The PD results indicate that, although the optimizer was formulated without explicit dependence on bacterial ecological composition, it implicitly favored sensitive dominance (N_S > N_R) across most patients. This outcome was achieved without any ecological constraints or modeling efforts, arising solely from the dynamic MIC-trait coupling (Eq. (4)), and represents a remarkable emergent property consistent with the containment principle that maintaining sufficiently large sensitive populations suppresses resistance expansion and preserves antibiotic efficacy⁶². This emergent behavior is shown in Fig. 8a, with green and purple regions indicating dominance of the sensitive and resistant populations, respectively. Across the virtual cohort, most optimized therapies concluded within 6–8 cycles, with only a minority extending beyond this range. These longer regimens, classified as extended and defined empirically as those exceeding 8 cycles, exhibited a consistent pattern of late-cycle bacterial rebound (Fig. 8a.1, b).

Fig. 8 — a Dominance heatmap of the virtual cohort under iterative optimization, highlighting the subset of patients undergoing extended treatments (>8 cycles). **a.1** Dominance patterns of this extended-treatment subgroup before MPC refinement. **a.2** Dominance patterns of the same subgroup after MPC refinement, showing a marked reduction in resistant dominance events and improved long-term ecological control. The dashed line at the 8th cycle separates standard from extended regimens; because cycle indices are discrete, patients completing cycle 9 or beyond are classified as extended treatments. b Median relative change in bacterial load ( $Δ \log_{10}$ CFU relative to baseline) per treatment cycle before MPC refinement, stratified into normal (<8 cycles) and extended (>8) regimens. c Corresponding trajectories after MPC refinement, showing improved suppression in patients previously requiring extended treatment. Shaded areas in (b) and (c) represent interquartile ranges of the virtual cohort.

MPC refinement mitigates late-cycle bacterial rebound and improves long-horizon control

The bacterial rebound in extended treatments occurred despite local PK-PD target attainment, revealing a limitation of the purely iterative optimization strategy: by optimizing each cycle independently, the algorithm neglected cumulative eco-evolutionary feedbacks between cycles. To mitigate these long-term effects, we investigated an MPC refinement acting over H = 3 cycles on a finite dose grid [10, 25] mg/kg. The controller maintains the same targets and constraints as the iterative single-cycle optimizer but augments the objective with a bacterial reduction term, while propagating the MIC through the resistance-trait dynamics (Eq. (4)) to anticipate bacterial adaptation rather than merely react to it.

When applied to patients on extended regimens, the MPC refinement successfully converted late-cycle rebounds into evident bacterial decline with the interquartile region shifted toward more negative values (Fig. 8c) and a substantial reduction and delay in resistant-dominance events, confirming improved ecological control without violating PK-PD safety limits (Fig. 8a.2).

Concluding insights on cohort-level ecological control and future directions

To capture how dominance patterns evolve across the entire population, Fig. 9 reports the cycle-wise prevalence (%) of sensitive (N_S > N_R) and resistant (N_R > N_S) bacterial dominance, together with the number of patients still under treatment (purple line). This representation aggregates the ecological dynamics observed in individual heatmaps (Fig. 8a) and provides a quantitative measure of cohort-wide bacterial control over time. Under the iterative single-cycle optimization (row (1)), sensitive dominance initially prevails in nearly all patients, but the proportion of resistant-dominant infections increases steadily after cycle 6. At first glance, the apparent drop in resistant prevalence beyond cycle 10 might suggest recovery of control, yet this is misleading: the purple line reveals that very few patients remain under treatment at that point. Hence, the illusion of recovery actually reflects attrition of the patient pool: most short-regimen cases have already cleared the infection, leaving only a small subset of rebounding patients contributing to the late-cycle statistics. After MPC refinement (row (2)), resistant dominance no longer accumulates with cycle number, and the fraction of sensitive-dominant infections remains above 80 % throughout, even as the number of active patients declines. This indicates that MPC not only accelerates bacterial clearance but also reshapes the ecological trajectory, preventing the enrichment of resistant subpopulations seen under extended regimens. Finally, focusing exclusively on the extended-treatment subgroup re-optimized under MPC (row (3)), resistant dominance rarely exceeds 15%, and sensitive prevalence remains stable across all gestational-age strata, confirming that the MPC controller effectively stabilizes long-horizon eco-evolutionary dynamics even in the most difficult cases, anticipating resistance drift by modulating exposure before it becomes selective.

DT-optimized dosing achieves improved bacterial control with clinically plausible schedules in real-world cohorts

To evaluate the clinical translatability of the evolutionary digital-twin optimizer, we applied the iterative single-cycle optimization (O) and classical (C) dosing schedules to two real-world neonatal cohorts, AMICREA-CI and AMICREA-TH, enabling an external assessment of the framework beyond the virtual population. In these cohorts, information on the relative abundance of susceptible and resistant subpopulations is not available, nor is information on the total bacterial burden from which such abundances could be inferred. Consequently, patient-specific initial conditions could not be defined. PD model calibration and subsequent simulations were therefore initialized using the same bacterial loads adopted in the virtual-population analysis, namely N_S(0) = 8.8 × 10⁵ and N_R(0) = 2.2 × 10⁵.

Table 5 reports two key outcome metrics across GA subgroups: (i) the Mean Final Resistance (MFR), defined as the mean log₁₀N_R measured at the end of treatment across all patients in each group; and (ii) the mean final reduction in total bacterial burden (MΔCFU), quantified as the mean Δlog₁₀(N_R + N_S) at the end of treatment relative to baseline.

Table 5.

Comparison between classical (C) and optimized (O) dosing evaluated on the real cohorts AMICREA-CI and AMICREA-TH

Real cohorts	GA (wk)	Samples (n)	MFR_C	MFR_O	MΔCFU_C	MΔCFU_O
AMICREA-CI	<28	60	7.7094	−6.8839	0.37168	−3.2785
	28–31	165	6.0586	−4.6612	0.10207	−2.9469
	32–34	124	5.2124	−2.9731	−0.05140	−2.7170
	35–37	133	5.6072	−0.9148	0.01595	−2.3918
	>37	226	5.0346	0.5280	0.11499	−1.7719
AMICREA-TH	34–36	10	5.6229	2.3305	0.18417	−1.9415
	≥37	46	5.7851	0.75856	0.25260	−1.9203

Open in a new tab

MFR Mean Final Resistance (mean log₁₀N_R at the end of each treatment), MΔCFU mean reduction in final total bacterial load (Δlog₁₀(N_R + N_S) relative to baseline).

Across all GA ranges and in both cohorts, the optimized regimens consistently achieve stronger bacterial clearance and markedly superior resistance suppression. Classical schedules show persistent or expanding resistant subpopulations (positive MFR_C), whereas optimized dosing drives MFR_O strongly negative, indicating near-elimination of resistant bacteria even in extremely and very preterm neonates. Similarly, bacterial burden reduction under optimized treatments (MΔCFU_O) improves by 2–3 log₁₀ units compared with classical dosing (MΔCFU_C).

Consistently with these metrics, Fig. 10 summarizes PK/PD target attainment at the last dosing cycle across MIC strata in the two cohorts.

Across both AMICREA-CI and AMICREA-TH, the optimized schedules preserve the key PK/PD goals, with universal attainment of $T_{C_{\min}} \leq 48$ h and $C_{\max} \geq 15$ mg/L across all MICs, and high AUC/MIC coverage for MIC ≤ 6 mg/L. At the same time, exposure remains largely within the predefined safety window, with limited proportions of patients crossing the $C_{\max} \geq 35$ mg/L soft cap.

Finally, to assess the clinical plausibility of DT-optimized regimens beyond PK/PD performance alone, administered and optimized schedules in the AMICREA-CI and AMICREA-TH cohorts were compared using (i) dosing-occasion-level metrics on dose amount (%Δ_dose) and inter-dose interval (Δτ) and (ii) a patient-level metric of cumulative exposure (%Δ_cum). Dispersion of relative metrics was summarized using the median absolute deviation (MAD). All these metrics are formally defined in the Methods. Comparisons were restricted to identical clinical decision points defined at the individual-neonate level by PNA, by considering (PtID, PNA) as joint key, and ensuring that optimized and administered regimens were evaluated under comparable physiological and maturational conditions. This yielded 76 and 32 one-to-one matched dosing occasions in the AMICREA-CI and AMICREA-TH cohorts, respectively. Results are summarized in Table 6.

Table 6.

Structural comparison between administered and DT-optimized regimens

Metric	AMICREA-CI	AMICREA-TH
Matched dosing occasions, n	76	32
Neonates contributing, m	37	14
Dose amount (occasion-level)
Median %Δ_dose	5.8%	47.7%
MAD %Δ_dose	23.0%	30.3%
Timing (inter-dose intervals, h; occasion-level)
Median Δτ	−6.3	−15.1
Mean Δτ ± SD	−10.0 ± 11.3	−20.1 ± 16.6
Cumulative exposure (patient-level)
Median %Δ_cum	3.5%	35.6%
MAD %Δ_cum	16.6%	26.4%

Open in a new tab

For dose amount and interval metrics, median, MAD, and mean± SD were computed across all occasion-level values {%Δ_dose,i} and {Δτ_i}, with i = 1, …, n. For cumulative exposure, median and MAD were computed across patient-level values {%Δ_cum,k}, with k = 1, …, m.

In AMICREA-CI, DT-optimized dose amounts closely aligned with the history of doses administered in the clinical setting. The distribution of %Δ_dose showed a near-zero central tendency (median +5.8%), with upward and downward adjustments occurring in a balanced manner across matched occasions (MAD 23%), indicating that optimization primarily refined, rather than systematically altered, existing dosing decisions. In contrast, AMICREA-TH exhibited a consistent upward shift (median %Δ_dose = + 47.7%) with greater dispersion (MAD 30.3%). In AMICREA-CI, the median %Δ_cum was +3.5%, indicating that improved bacterial control was achieved without increasing overall exposure, but through redistribution of dose timing within standard schedules. Conversely, AMICREA-TH showed higher cumulative exposure (median %Δ_cum = + 35.6%), accompanied by systematically shorter inter-dose intervals (median reduction of ~15 h).

Overall, these results demonstrate that DT-optimized regimens substantially improve bacterial clearance and resistance control in real-world neonatal cohorts without sacrificing PK safety and while remaining clinically plausible. In AMICREA-CI cohort, these gains are achieved through refined redistribution of dosing over time without increasing cumulative exposure, whereas in AMICREA-TH slightly more pronounced differences in dosing patterns are observed, reflecting the documented tendency of the model to underpredict exposure, prompting the optimizer to increase dose amount and dosing frequency to ensure robust attainment of the specified PK/PD targets.

Although the limited number of matched occasions, especially in AMICREA-TH, influences the precision of these estimates, the direction and magnitude of the observed differences were consistent with the patterns emerging from the virtual-cohort analysis (Supplementary Note 4), where DT-optimized schedules exhibited coherent trends across maturational and clinical factors, supporting the translational potential, clinical feasibility, and generalizability of the DT framework across heterogeneous neonatal populations.

Discussion

The present study establishes an evolutionarily-informed DT framework for optimizing individual aminoglycoside therapy in neonates with suspected sepsis, using amikacin as a case study. By integrating real-world neonatal data with a PBPK-PD model and a data-driven renal function predictor, we provide a unified view on how neonatal physiology and bacterial evolution jointly shape antibiotic efficacy and emerging resistance. This system-level approach enables adaptive, individualized therapeutic control under clinically realistic conditions, moving beyond conventional dosing paradigms that rely on average, static disease profiles and overlook dynamic patient heterogeneity. The value of integrating mechanistic knowledge with data-driven techniques was confirmed through extensive simulations demonstrating consistent improvements in joint PK-PD indices of therapeutic success.

In detail, the data-driven layer, implemented as an LSTM-based neural network, continuously estimates eGFR and its alteration under NICU-relevant conditions such as AKI and PATH, achieving a mean MAPE reduction of 29% on average (range: 8–44%) in benchmark experiments against a mechanistic kidney model for sCr prediction¹⁸. The increased accuracy of sCr predictions achieved with an LSTM network further refined amikacin PK modeling across therapy, highlighting the reliability of the PBPK model’s dynamic updating. These continuous tracking predictions dynamically inform the PBPK model, deliberately restricted to the compartments essential for amikacin kinetics (kidney, heart, serum blood, and remainder), so that it can be parameterized using standard, low-dimensional, routinely available clinical data with no additional costs or patient burden. We applied the framework to both real and virtual patients; the resulting simulated trajectories enabled continuous assessment of neonatal conditions by faithfully reproducing the recorded amikacin concentration profiles and blood culture-proven outcomes.

Although the PD module introduces additional parameters to describe bacterial ecology and resistance evolution, these were carefully calibrated to connect dosing regimens with their ecological and evolutionary outcomes. Related results unleash new connection between standard regimen and their PD effects on both ecological and evolutionary perspective, thus providing a comprehensive new model to accurately inform precision dosing. Together, these elements establish a solid framework for therapy optimization, enabling continuous risk assessment and personalized dosing strategies that respect pharmacological and safety limits. In this context, the evolution-aware DT framework demonstrated excellent performance in identifying patient-specific, safe, and effective regimens across the simulated neonatal population. These in silico results, grounded in retrospective clinical data, demonstrate the feasibility of individualized dosing under clinically realistic conditions and provide a strong quantitative rationale for prospective evaluation of clinical outcomes, while acknowledging that framework applicability should be interpreted within the range of variability captured by the real-world cohorts and reliably explored in the virtual population.

A key feature of this eco-evolutionary design is the dynamic MIC, which evolves according to resistance-trait dynamics, allowing the optimizer to anticipate bacterial adaptation rather than reacting to it. During each optimization cycle, dosing decisions targeted a peak-to-MIC ratio of $C_{\max} / M I C \geq 8$ , an interdose interval of $T_{C_{\min}} \leq 48$ h, and exposure safety thresholds of $C_{\max} \leq 60$ mg/L (penalized above 35 mg/L). These criteria ensured 100% compliance with the interdose constraint and full attainment of bacteriostatic exposure (AUC/MIC ≥ 21.4) across all MIC strata. Near-complete bactericidal coverage (AUC/MIC ≥ 62.5) was achieved in 94–98% of neonates with MIC ≤ 6 mg/L, and 58% for MIC 6–8 mg/L, reflecting pharmacological rather than algorithmic limits. The optimizer consistently maintained $C_{\max} / M I C \geq 8$ for MIC ≤ 6 mg/L, while safely capping exposures for higher MICs without compromising efficacy. Note that $T_{C_{\min}}$ was defined as the first post-infusion time when serum concentration fell below a conservative $C_{\min} = 3$ mg/L, under which excellent PK/PD performance was already achieved, suggesting that a relaxed threshold (e.g., 5 mg/L²⁴) could further enhance coverage. Importantly, for patients whose $C_{\max} / M I C$ ratio fell below 8, primarily those with baseline MIC values between 6 and 8 mg/L, this condition does not imply an increased risk of selecting for resistance. As reviewed by Andersson and Hughes⁷, resistant mutants are typically enriched only at highly subinhibitory antibiotic levels, i.e., concentrations several hundred-fold below the MIC, far below the ranges achieved in our dosing scenarios.

In a subset of extended treatments, bacterial rebound emerged from cumulative eco-evolutionary feedback across dosing cycles. Thus, we implemented an MPC strategy to anticipate resistance drift, suppressing late-cycle rebounds and stabilizing long-term ecological dynamics while preserving pharmacological safety. Related results indicate that MPC refinement not only shortens therapy but also rebalances bacterial ecology by sustaining sensitive dominance through anticipatory control of the MIC trajectory, unveiling a promising direction for future research.

Our modeling framework, due to its modular development, can be extended to new scenarios involving combinatorial antimicrobial regimens and gradually increasing resistance risk, by refining PD parameters based on expected drug concentration-time profiles. Through close collaboration between neonatologists and modeling scientists, the parameterization was conceived to remain both feasible and clinically interpretable, linking mathematical abstraction to measurable clinical data⁶³. Although amikacin is the only aminoglycoside explicitly instantiated and validated in this study, due to the availability of rich neonatal TDM data and its widespread use in suspected neonatal sepsis, the proposed digital-twin framework is not inherently drug-specific. Its modular structure separates the PBPK layer, the renal function predictor, and the eco-evolutionary PD module, enabling substitution of compound-specific parameters without altering the overall control architecture. Extension to other aminoglycosides would just require re-parameterization of drug-specific PK parameters, PD potency terms, and optimization target ranges, while preserving the overall model structure and control strategy. Importantly, the PD parameter ranges used in this work were informed by trends reported in the literature for gentamicin and subsequently rescaled to account for the lower intrinsic potency of amikacin. As a result, adaptation of the framework to gentamicin would require limited re-scaling of potency-related parameters rather than structural model changes.

Although our case study focused on amikacin, an antibiotic for which resistance mechanisms are relatively less prevalent than for other antimicrobial agents, in silico simulations in large virtual populations showed that prolonged antibiotic therapy may increase the risk of bacterial rebound, a phenomenon consistent with observations in the microbiology literature⁶⁴. This behavior further supports the integration of our evolutionary dosing framework in the design of combination therapies and antibiotic cycling strategies aimed at modulating bacterial resistance and leveraging collateral sensitivity, in which resistance to one antibiotic increases vulnerability to another⁶⁵.

In terms of clinical scope, the proposed framework supports dosing optimization in suspected neonatal sepsis, encompassing both microbiologically confirmed, culture-negative cases, as well as patients who become culture-negative or culture-positive during therapy. A key advantage is its compatibility with empirical antibiotic use in the NICU, where culture confirmation and repeated sampling are often unavailable. In the absence of microbiological data, the eco-evolutionary component is initialized using standardized, empirically derived estimates; however, dosing optimization relies on relative bacterial dynamics rather than absolute burden, making the framework robust to uncertainty in the assumed initial bacterial condition and broadly applicable across real-world clinical scenarios. This design enables clinical translation within existing NICU workflows, using routinely available data and without requiring additional invasive procedures.

Three aspects merit further investigation. First, the current PBPK structure, while well suited for aminoglycosides, may require adaptation for antibiotics with different susceptibility to resistance development, e.g., gentamicin, as well as pharmacokinetic behavior, such as β-lactams or macrolides. Additional modifications may also be necessary for infection sites where drug penetration and local bacterial dynamics differ substantially from systemic circulation. Second, although the MPC refinement proved effective in simulation, its translation to real-time clinical decision support will require prospective validation, including uncertainty quantification and response to incomplete or delayed clinical data. Finally, extending the framework to special neonatal subgroups remains an important direction. In addition to oncologic patients, in whom chemotherapy and polypharmacy markedly alter aminoglycoside kinetics and toxicity, other clinically relevant factors, such as exposure to concomitant nephroactive or clearance-modifying medications and extracorporeal support modalities including Extracorporeal Membrane Oxygenation, may further influence aminoglycoside pharmacokinetics and were not explicitly modeled in the present study. Addressing these sources of variability and expanding the representation of extremely preterm neonates through larger dedicated cohorts will be important to further strengthen the translational relevance and generalizability of the proposed framework.

Methods

Data

Three real-world cohorts of neonates with suspected sepsis or severe infections admitted to the neonatal intensive care unit (NICU) of the University Hospitals Leuven (Belgium) were analyzed. Ethics approval for this study was obtained from the Ethics Committee Research (EC Research), UZ/KU Leuven, Belgium (reference S70058, approval date March 14, 2025). All research procedures were conducted in accordance with the principles of the Declaration of Helsinki, ICH-GCP, and the Oviedo Convention on Human Rights and Biomedicine. All data were anonymized to protect patient privacy, and informed consent was waived, as the study involved retrospective analysis of existing data without any direct patient interaction.

Each dataset reports key baseline maturational parameters, including birth weight (BW), gestational age (GA), and postnatal age (PNA) alongside clinical variables capturing additional comorbidities, such as acute kidney injury (AKI), perinatal asphyxia, and concomitant interventions, e.g., therapeutic hypothermia (TH), respiration supports and ventilation, or co-administration with ibuprofen and/or inotrope (CI).

Amikacin predictive models were built and evaluated on two longitudinal TDM datasets, AMICREA-TH and AMICREA-CI, which include full dosing histories (dose, administration time, infusion rate, weight-normalized dose) with paired plasma amikacin concentrations (mg/L) and sCr measurements (mg/dL). Lastly, the third dataset contains sCr measurements from patients with AKI and TH, namely CREA-AKITH. An overview of the clinicopathological features in each dataset is provided in Table 7. To compare the distribution of continuous variables, we used the two-sided Wilcoxon-Mann-Whitney test and reported the relative p-value.

AMICREA-CI dataset⁶⁶ includes a total of N = 709 patients with 3573 TDM samples; GA ranged from 24 to 42 weeks (median ≈34 weeks), with BW spanning 385–4650 g (median 2102.5 g). The median PNA and postmenstrual age (PMA) were 2 days and 34 weeks, respectively. Each patient had longitudinal treatment data paired with blood culture results, enabling continuous monitoring of bacterial infection over the course of therapy. The prevalence of Small for Gestational Age, defined as BW below the expected percentile for their GA was ~9%. Ventilation and additional respiratory support (e.g., Continuous Positive Airway Pressure) were required in ≈44% and 62% of patients, while inotrope and ibuprofen co-administration occurred in 19% and 7% of cases, respectively. Per-patient GA stratification showed that ibuprofen co-administration occurred in 26.7% of extremely preterm neonates (GA < 28 weeks), 12.1% of very preterm (28–32 weeks), 3.5% of moderate-to-late preterm (32–37 weeks), and 0.9% of term (≥37 weeks) infants. Inotrope use followed a similar trend, with 55.0%, 27.9%, 10.1%, and 12.8% across the respective GA categories, reflecting the greater hemodynamic support required in the most immature neonates to maintain adequate systemic perfusion.

AMICREA-TH²⁵ includes a total of 158 TDM samples collected from N = 56 neonates with perinatal asphyxia treated with whole body TH (PATH). GA ranged from 35 to 41 weeks (median 38 weeks), with BW spanning from 1910 to 4770 g (median 3037.5 g). The median PNA at sampling time was 2 days, while the median PMA was 38.5 weeks. Perinatal asphyxia was documented in 100% of patients, with TH applied in 98% and ventilation required in 76.8% of cases, while 96.4% of patients received additional respiratory support, and 55.3% inotrope co-administration. Ibuprofen co-treatment was not reported (0%). The gender distribution was 58% male and 41 % female.

CREA-AKITH⁵¹ includes a total of N = 869 neonates with 5444 sCr measurements collected in whole body TH-treated asphyxiated patients, some of whom also had AKI. As there were no data on urine output, the AKI definition was based on sCr trends, using the KDIGO definition as done in previous study⁶⁷. GA ranged from 34 to 43 weeks (median 40 weeks), with BW spanning 1750–6230 g (median 3335 g). The median PNA at sampling was 2.5 days. AKI, defined according to neonatal sCr criteria (increase ≥0.3 mg/dL within 48 h, or ≥1.5-fold from baseline within 7 days), was documented in ≈15% of patients. Stratification by GA revealed higher AKI prevalence in the most immature neonates, reflecting reduced renal reserve and greater vulnerability to perinatal asphyxia. Notably, mortality was markedly increased in neonates with AKI compared to those without, consistent with previous clinical evidence⁵⁴.

PBPK-PD model structure

The mechanistic ODEs specifying the PBPK-PD model were developed using the SimBiology toolbox v23.2 in MATLAB^® (Natick, MA, USA) for downstream simulation of the interactions with infection control actions.

Figure 11 illustrates the PBPK-PD model together with the governing equations of the PBPK module, adapted from the physiologically based formulation in Zazo et al.⁶⁸, which describes amikacin distribution and elimination across four physiological compartments: heart, kidneys, remainder tissues, and serum (blood), whose drug concentrations are denoted by C_h, C_k, C_r and C_b, respectively. The dose (A_in) is infused into the serum blood compartment. The PD module was embedded within the same compartment, capturing systemic inflammatory responses characteristic of sepsis, and was directly coupled to the PBPK model through the time-varying serum amikacin concentration C_b, which defined the selective pressure on bacteria strains. Two bacterial subpopulations were modeled as SimBiology species: drug-susceptible (N_S) and drug-resistant (N_R) bacteria. Their eco-evolutionary competition under antibiotic exposure was described by the system of ODEs already presented in the Results, where the resistant subpopulation features a time-varying adaptive trait u that modulates drug susceptibility.

Fig. 11 — Left Schematic of the PBPK-PD system describing amikacin infusion, distribution and elimination across four physiological compartments: heart, kidneys, remainder tissues and serum blood, where drug is administered (A_in). The PD module captures the eco-evolutionary competition between drug-susceptible (N_S) and drug-resistant (N_R) bacterial strains in the blood. Green nodes denote model species, yellow circles represent reactions. The adaptive trait u (orange node) represents a time-varying parameter modulating bacterial antibiotic resistance. The purple dashed arrow indicates the antibiotic selective pressure exerted through C_b on bacterial fitness. Right System of ODEs defining the PBPK component, including concentrations (C_*, mg/L), inter-compartmental flows (Q_*, L/h), volumes (V_*, L), and dimensionless partition coefficients (P_*), together with infusion dynamics and systemic clearance. Parts of this figure were created with BioRender.com.

Amikacin is predominantly eliminated by glomerular filtration. In the PBPK model, renal clearance (CL) is implemented as a first-order removal term proportional to circulating drug concentration C_b, assuming perfusion-limited distribution, and representing renal excretion as a lumped systemic clearance pathway rather than an explicit nephron-level process. Furthermore, no absorption compartment was included, as amikacin is administered intravenously. The administered amount A_in, given by the product of the optimized dose (mg/kg) and current neonatal body weight (CW, in kg), enters the central compartment through a zero-order infusion process. The corresponding infusion rate R is given by $A_{i n} / t_{\inf}$ during the infusion interval $[0, t_{\inf}]$ and is zero thereafter, where $t_{\inf}$ denotes the infusion duration. Drug distribution is governed by organ blood flow rates (Q_i, with i = h, k, r) and tissue to blood partition coefficients (P_i, with i = h, k, r). Blood flow rates (Q_h, Q_k, Q_r) were scaled to each subject’s predicted CO using fixed fractional distributions of 5% (heart), 20% (kidneys), and 75% (remainder).

Compartmental volumes were individualized from CW using fixed fractions derived from neonatal anatomical data⁶⁹. The blood compartment volume (V_b) was computed from CW scaled by a volume coefficient for blood (BV)⁷⁰, as follows:

V_{b} = C W_{kg} \cdot B V_{perkg}, B V_{perkg} = \{\begin{matrix} 0.106, if GA < 37 weeks (preterm), \\ 0.080, if GA \geq 37 weeks (term) . \end{matrix})

Organ volumes were set as fixed fractions of CW: 0.7% for the heart and 1.0% for the kidneys⁷¹, with the remainder compartment defined by difference to ensure total mass balance.

The partition coefficient of the remainder compartment (P_r) was adjusted to reproduce the observed apparent volume of distribution of amikacin (V_d,target)⁷², using:

P_{r} = \max (\frac{V_{d, target} - V_{b} - P_{h} V_{h} - P_{k} V_{k}}{V_{r}}, 0.05) .

PD model calibration procedure

For each neonate, we identified the eco-evolutionary PD parameter vector

θ = [E C 5 0_{S_{0}}, E C 5 0_{R_{0}}, E_{\max}, c, r_{S}, r_{R}, k_{u}, σ, d] \in R_{\geq 0}^{9},

where $R_{\geq 0}$ denotes the set of non-negative real numbers, to ensure that the coupled PBPK-PD system reproduced qualitatively plausible bacterial dynamics while referencing standard neonatal aminoglycoside dosing. In the proposed framework, PD parameters cannot be uniquely identified from clinical data alone, since the retrospective neonatal cohorts do not include longitudinal bacterial burden measurements. Accordingly, direct fitting of eco-evolutionary PD parameters to clinical microbiological outcomes was not feasible.

To address this limitation, we adopted a literature-informed strategy using the gentamicin PD curves reported in ref. ⁴¹ as an external behavioral reference, which provides characteristic bacterial killing-regrowth dynamics for a limited set of discrete GA groups (25, 29, 34, and 40 week). To extend the PD calibration across the full GA range (24–42 weeks), despite the availability of bacterial time-kill reference curves only at discrete intermediate GAs, we designed a two-steps, constrained calibration procedure. First, we subdivided the full GA range into four intervals, centered on the available anchors, i.e., [24–27], [28–32], [33–37], [38–42]. We adapted the corresponding gentamicin exposure profiles to amikacin by accounting for its lower intrinsic potency⁷³. Subsequently, the available reference time-kill curves (corresponding to representative GA anchors) were fitted using empirically initialized PD parameters to reproduce the aggregate dynamics of the total bacterial population (N_S + N_R). We considered admissible all the PD curves lying in the neighborhood of each GA-specific reference curve, which were interpreted as representing expected inter-patient variability in the relative GA sub-range. From this set of accepted trajectories, the admissible parameter range vector $\tilde{Θ}$ was derived as the component-wise minimum and maximum values attained by each PD parameter. These parameter ranges are reported in Table 1.

Because assignment of an individual patient to a reference GA cluster and its associated dynamics is an approximation, the resulting parameterizations were not assumed to be directly transferable to individual patients. Therefore, in the second calibration step, a patient-specific parameterization was performed to identify an individual parameter vector θ(5) for each neonate, constrained to lie within the admissible range $\tilde{Θ}$ (Table 1). This estimation relied on individual PBPK-derived amikacin exposure profiles obtained under standard-of-care dosing based on CW and PNA, as reported in ref. ²⁴. Dosing schedules were defined by the administered dose (mg/kg), the dosing interval τ, and infusion duration ( $t_{\inf} = 1$ h). τ ranged from 20 h in term neonates (CW > 2800 g) to 48 h in extremely preterm infants (CW < 800 g), consistent with clinical recommendations²⁴. Simulations were performed over $t_{\max} = \max (72 h, 3 τ)$ to capture multiple dosing cycles.

Individual model calibration was formulated as a constrained optimization problem:

\min_{θ \in \tilde{Θ}} ∥ r (θ) ∥_{2}^{2},

where the residual vector r(θ) aggregated constraint violations quantifying deviations from mechanistically expected PBPK-PD dynamics. Each term implemented a guard condition that constrained the simulated bacterial dynamics across dosing schedules to remain mechanistically realistic by limiting total bacterial burden, preventing premature extinction of susceptible cells, bounding resistant expansion, maintaining susceptible dominance early in treatment, constraining early-kill slopes, and ensuring a transient resistant advantage during antibiotic exposure with recovery of susceptibility after clearance. A detailed description of the penalty terms is provided in Supplementary Note 2.

Minimization was performed using a damped least-squares (Levenberg–Marquardt) algorithm (lsqnonlin, MATLAB^®) with tolerances of 10⁻⁸ and limits of 150 iterations and 1500 function evaluations. The coupled PBPK-PD system was numerically integrated using the ode45 solver in MATLAB^® (RelTol = 10⁻⁶, AbsTol = 10⁻⁹). Infusions were represented as zero-order inputs applied at intervals τ. Initial bacterial loads were set to 4.19 × 10⁵ × s_N and 1.26 × 10⁵ × s_N for N_S and N_R subpopulations, respectively, where s_N = 10³ serves as a scale factor for simulations expressed in CFU/L. These values correspond to a total baseline inoculum N₀ ≈ 5 × 10⁵ CFU/mL, in agreement with neonatal time-kill data⁴¹, and the bacterial dynamics were simulated to mimic those observed in serial follow-up blood culture experiments.

LSTM modeling of serum creatinine dynamics

In each dataset, multiple sCr measurements were available per patient, collected at varying time points and in different number across individuals. To reconstruct renal function dynamics influenced by both maturational and non-maturational factors, we trained LSTM deep neural networks to predict sCr values and subsequently compute eGFR, a key parameter in the PBPK model, using the Schwartz equation⁷⁴. The datasets were first preprocessed to remove outliers and ensure data consistency. Only records with positive sCr values ≤2 mg/dL in the AMICREA-TH and AMICREA-CI, and ≤4 mg/dL in the CREA-AKITH cohorts were retained, and any measurements row containing missing values in key variables (e.g., GA, PNA, and sCr) was removed. This step ensured that the model received complete and reliable input data, without introducing bias through imputation. It is worth noting that only few samples were discarded in each cohort, and the following data were used downstream: 155 TDM measurements from 55 patients in the AMICREA-TH cohort, 3491 TDM measurements from 699 patients in the AMICREA-CI cohort, and 5443 sCr measurements from 869 patients in the CREA-AKITH cohort.

To account for irregular sampling inherent in clinical time-series, all data were structured on a per-patient sequence basis, where each sequence contains temporal measurements of relevant covariates (PNA, GA, body weight, AKI status, TH, inotropic and ibuprofen therapy) and corresponding sCr values. Sequences were sorted by length prior to batching to minimize the amount of zero-padding required when aligning sequences within each mini-batch. During training, a mini-batch size of 16 was selected as a trade-off between reducing variability in sequence lengths within a batch and ensuring training stability and convergence. Within each mini-batch, shorter sequences were padded only up to the length of the longest sequence in that batch, thereby limiting the proportion of padded values processed by the network, improving computational efficiency, and ensuring that learning is driven primarily by observed clinical measurements rather than artificial padding.

To prevent information leakage, all repeated measurements from the same patient were consistently allocated to either the training or the test set, and a stratified k-fold cross-validation was employed, ensuring balanced patient representation across folds according to GA and PNA strata. No data normalization was applied. Each neonate in AMICREA-CI dataset was encoded by the following feature vector: {GA, PNA, CW, inotrope, ibuprofen}. The CREA-AKITH model for neonates with AKI employed {GA, PNA, AKI}. Both workflows shared the same deep neural network architecture, consisting of three stacked LSTM layers with 500, 250, and 100 units, respectively. The hyperbolic tangent (tanh) function was employed as the state activation, while the sigmoid function (sig) served as the gating mechanism. To mitigate overfitting, dropout layers (10%), which helps regularize training by randomly masking a fraction of neurons during each training step update, were inserted between consecutive LSTM layers. Finally, a fully connected layer feeds into the regression output layer, optimized by minimizing the mean squared error loss function. For neonates with PATH in AMICREA-TH cohort, the input feature vector was {GA, PNA, TH}. Due to the smaller number of available observations, a smaller deep learning architecture with a lower number of learnable layers, consisting of only two LSTM layers containing 400 and 200 cells, followed by a dropout layer (10%), and a final fully connected regression layer for predicting continuous sCr values was trained. Each cell of LSTM consists of three inputs: previous cell state, previous hidden state, and input for current time step and provides next layer with two outputs: current cell state and current hidden state. The three gates are: forget gate, determining which information is to be erased or forgotten from the memory unit, the input gate determining what information should be added in the long-term memory or cell state, and finally the output gate to decide what to give as output from the long-term memory to the following time step. Figure 12 provides a graphical overview of the LSTM structure, including the definitions of its intermediate signals and governing equations.

Fig. 12 — Left LSTM architecture, implemented as a gated recurrent unit with forget, input, and output gates regulating information flow through the cell state (C_t) and hidden state (h_t). Right Recurrent update equations defining the LSTM model, where gating variables (f_t, i_t, o_t) control the information retained, added, or output at each time step. $x_{t} \in R^{p \times m}$ is the input matrix containing m features observed over p consecutive time points; $h_{t} \in R^{n_{h}}$ represents the hidden state vector at time t; $C_{t} \in R^{n_{h}}$ is the cell state vector (memory); and ${\hat{y}}_{t} \in R$ is the predicted sCr at time t. U_*, W_*, and V are learnable weight matrices, and β_*, β^(y) are bias terms. g( ⋅ ) is the identity function for sequence-to-one regression task. Parts of this figure were created with BioRender.com.

Model performance metrics

The following definitions of the mean prediction error (MPE) and the mean absolute prediction error (MAPE) were respectively used to assess prediction bias and to estimate prediction accuracy of sCr prediction through LSTM models for a cohort of N patients, each having n_i sCr measurements. Mathematically, they are defined as:

M P E = \frac{1}{N} \sum_{i = 1}^{N} \frac{1}{n_{i}} \sum_{j = 1}^{n_{i}} (\frac{{sCr}_{{pred}_{i j}} - {sCr}_{{obs}_{i j}}}{{sCr}_{{obs}_{i j}}}), M A P E = \frac{1}{N} \sum_{i = 1}^{N} \frac{1}{n_{i}} \sum_{j = 1}^{n_{i}} \frac{|{sCr}_{{pred}_{i j}} - {sCr}_{{obs}_{i j}}|}{s C r_{o b s_{i j}}}

The Average Fold Error (AFE) and Mean Absolute Error (MAE) were used to assess the accuracy of the PBPK model in predicting n_i amikacin concentration measurements after dosing, and are defined as:

A F E_{i} = 1 0^{\frac{1}{n_{i}} \sum_{j = 1}^{n_{i}} \log_{10} (\frac{{sCr}_{{pred}_{i j}}}{{sCr}_{{obs}_{i j}}})}, M A E_{i} = \frac{1}{n_{i}} \sum_{j = 1}^{n_{i}} ∣C_{p r e d_{i j}} - C_{o b s_{i j}}∣

where ${s C r}_{{p r e d}_{i j}}$ is the predicted concentration for the j-th observation, ${s C r}_{{o b s}_{i j}}$ is the corresponding observed concentration, and n_i is the total number of data points for the i-th patient.

Optimization and simulation

Therapy optimization was formulated as a constrained nonlinear minimization problem. The PBPK and eco-evolutionary PD components define the dynamic system on which the optimizer operates. Physiological parameters are continuously updated through a combination of physiologically based equations (Supplementary Note 3) and the creatinine-driven eGFR model. Remarkably, the LSTM model is selected at each iteration in accordance with the patient’s evolving pathophysiology. For example, the TH remains active during three days, after which the system transitions to the appropriate patient-state model.

For each virtual neonate i, at each dosing cycle k, the optimizer determined the individual dose D_i,k by solving an optimization problem over a finite sequence of dosing cycles, subject to mechanistic and safety constraints. The objective function penalizes the deviations of simulated PBPK-PD dynamics, driven by candidate dose D_i,k, from predefined efficacy ( $C_{\max} / M I C \geq 8$ ), feasibility ( $T_{C \min} \leq 48$ h), and safety ( $C_{\max} \leq 35$ mg/L, hard-capped at 60 mg/L) targets, leading to the following expression for the running cost at each dosing cycle:

J_{i, k} (D_{i, k}) = w_{e ff} {(8 - C_{\max_{i, k}} / {M I C}_{i, k})}_{+}^{2} + w_{i n t} {(T_{C \min_{i, k}} - 48)}_{+}^{2} + w_{t o x} {(C_{\max_{i, k}} - 35)}_{+}^{2},

where ${(x)}_{+} = \max (0, x)$ .

The explicit functional dependence of $C_{\max, i, k}$ , $T_{C \min, i, k}$ , and MIC_i,k on the dose D_i,k is omitted for readability but remains implicit in all expressions. The feasible dose domain was D_i,k ∈ [10, 25] mg/kg, administered as a 1-h constant intravenous infusion ( $t_{\inf} = 1$ h). For each individual, the PBPK input amount was then computed as A_in = D_i,k × CW_i,k, and supplied to the central serum blood compartment through the infusion rate

R_{i, k} (t) = \{\begin{matrix} \begin{matrix} A_{in} / t_{\inf}, & 0 \leq t \leq t_{\inf}, \end{matrix} \\ t_{\inf} = 1 h . \\ \begin{matrix} 0,, t > t_{\inf}, \end{matrix} \end{matrix})

The terminal PK state from cycle k defines the initial condition for the next one, thus closing the adaptive feedback loop. Furthermore, for each neonate i, at each cycle k, physiology and bacterial susceptibility are consistently updated as described below.

sCr_i,k is estimated from the state-conditional LSTM-based neural network model. In detail, sCr predictions are generated by using the model trained on the CREA-AKITH cohort if the neonate has AKI and undergoes TH, by using the model trained on the AMICREA-TH cohort if is undergoing TH, and the model trained on the AMICREA-CI cohort, otherwise. The updated sCr value yield individual cycle-specific eGFR estimate (eGFR_i,k) via the Schwartz equation⁷⁴:
$e {GFR}_{i, k} = \frac{k {Ht}_{i, k}}{s {Cr}_{i, k}}$
where k is a GA-dependent Schwartz constant (set to 0.33 if GA < 37 weeks or to 0.45 otherwise). Finally, the clearance (CL_i,k), in L/h, is computed as:
${CL}_{i, k} = {eGFR}_{i, k} \frac{{BSA}_{i, k}}{1.73},$
BSA_i,k, CW_i,k Ht_i,k are updated through PNA-dependent functions (Supplementary Note 3). Analogously, CO, tissue blood flows, and volume of distribution are updated by using GA-, PNA-, sex and clinical information (TH, inotropes, ibuprofen, and PDA) as covariates. Collectively, the resulting physiological state serves as the individualized PBPK setup for the following cycle.
The bacterial resistance trait u evolves according to Eq. (3), and the corresponding effective MIC follows a saturating dynamic relationship, reported in Eq. (4).

Model-predictive control formulation

To mitigate late-cycle rebounds observed under extended regimens, a Model Predictive Control (MPC) problem was formulated over a finite sequence of dosing cycles. Let the therapy be divided into discrete decision instants indexed by $t \in I = {1, 2, \dots, N_{c y c}}$ , where each index t corresponds to the start of one dosing cycle (including drug administration and subsequent 72-h post-infusion simulation window). At each decision step t, the controller predicts system evolution over a finite prediction horizon of H = 3 cycles, defined as $T (t) = {t, t + 1, \dots, t + H - 1}$ .

The objective function in this formulation extends that in Eq. (8) by incorporating an additional penalty on insufficient bacterial reduction across cycles, quantified as the median relative change in bacterial load relative to baseline ( $Δ \log_{10} {C F U}_{i, k}$ ). In detail, the running cost at cycle k for patient i becomes:

{\tilde{J}}_{i, k} (D_{i, k}) = w_{e ff} {(8 - C_{\max_{i, k}} / {M I C}_{i, k})}_{+}^{2} + w_{i n t} {(T_{C \min_{i, k}} - 48)}_{+}^{2}

+ w_{t o x} {(C_{\max_{i, k}} - 35)}_{+}^{2} + w_{d} {(Δ \log_{10} {C F U}_{i, k} + ε_{k i l l})}_{+}^{2}

As in Eq. (8), the dose dependence is implicit and omitted for readability.

The optimal dose sequence is computed as:

\begin{matrix} {{\tilde{D}}_{i, k}^{t}}_{k = t}^{t + H - 1} = a r g \min_{{D_{i, k}}_{k = t}^{t + H - 1}} \sum_{k = t}^{t + H - 1} {\tilde{J}}_{i, k} (D_{i, k}) \\ s . t . {\overset{°}{x}}_{i, k} = f_{P B P K} - P D (x_{i, k}, D_{i, k}, p_{i}), x_{i, k} (0) = x_{i, k - 1} (T), \\ {M I C}_{i, k} = {M I C}_{i, k - 1} + ({M I C}_{\max} - {M I C}_{i, k - 1}) \frac{u_{i, k - 1}^{h}}{u_{50}^{h} + u_{i, k - 1}^{h}}, \\ 10 \leq D_{i, k} \leq 25 m g / k g, T_{C \min, i, k} \leq 48 h, C_{\max_{i, k}} \leq 60 m g / L, \\ N_{S_{i, k}}, N_{R_{i, k}} > 0 . \end{matrix}

Here, x_i,k is the PBPK-PD state vector at cycle k, including amikacin serum concentration, bacterial populations (N_S, N_R), and the adaptive trait u. The function f_PBPK−PD denotes the coupled mechanistic dynamics of the PBPK and eco-evolutionary PD layers and ${(x)}_{+} = \max (0, x)$ applies soft penalties to target violations. $T_{C \min_{i, k - 1}}$ is the previous cycle’s dosing interval. At each control step t, only the first action ${\hat{D}}_{i, t}$ from the optimized sequence computed at time t is administered:

\hat{D_{i, t}} : = {\tilde{D}}_{i, t}^{t},

after which the PBPK-PD model is re-simulated to obtain the updated state x_i,t(T). This state, along with the newly predicted renal clearance and evolved MIC, forms the initial condition for the next optimization at t + 1, thus implementing a standard receding-horizon feedback loop. By explicitly propagating MIC through the adaptive trait, the controller anticipates eco-evolutionary feedbacks, enabling smoother long-horizon dose adaptation and reducing late resistance rebounds.

Comparison metrics between real and optimized schedules

Comparisons between administered and optimized dosing schedules in the AMICREA-CI and AMICREA-TH cohorts were restricted to identical clinical decision points defined for each neonate by PNA, using (PtID, PNA) as a joint key to ensure that both regimens were evaluated under comparable physiological and maturational conditions.

For each matched occasion i within neonate k, relative dose differences were quantified as

% Δ_{d o s e, i} = 100 \times (\frac{{Dose}_{O, i}}{{Dose}_{C, i}} - 1),

where Dose_C,i denotes the administered dose (mg/kg) and Dose_O,i the corresponding DT-optimized dose (mg/kg). To characterize exposure over the entire matched observation window, a patient-level cumulative dose intensity was defined for each regimen j ∈ {O, C} as

I_{j, k} = \frac{\sum_{i \in I_{k}} {Dose}_{j, i}}{t_{last, k} - t_{fi r s t, k}}, [mg / kg / day],

where $I_{k}$ is the set of matched occasions for neonate k, Dose_j,i is the weight-normalized dose (mg/kg) for patient k at the i-th occasion, and t_first,k and t_last,k denote the times of the first and last matched occasions (expressed in days). Relative differences in cumulative exposure were then computed for each neonate as

% Δ_{c u m, k} = 100 \times (\frac{I_{O, k}}{I_{C, k}} - 1),

yielding one value per neonate. Because these relative metrics were ratio-based and exhibited skewed distributions, variability was summarized using the median absolute deviation (MAD), defined as

m e d i a n (∣ z - m e d i a n (z) ∣),

where z denotes the set of values under consideration. This metric was applied at two aggregation levels: (i) to the set {%Δ_dose,i} pooled across all matched occasions from all neonates; and (ii) to the set {%Δ_cum,k} across neonates.

Differences in dosing frequency were evaluated by comparing the inter-dose intervals of the optimized and administered schedules. For each matched occasion i, the interval change was defined as

Δ τ_{i} = τ_{O, i} - τ_{C, i},

where τ_C,i denotes the time from dose i to the subsequent administered dose, and τ_O,i the corresponding interval recommended by the DT optimizer. These differences were summarized across occasions using the mean± SD to describe directional shifts in dosing frequency and the median to quantify the typical magnitude of scheduling adjustments.

Supplementary information

Supplementary Information^{(2.7MB, pdf)}

Acknowledgements

This study was funded by: Italian Ministry of University and Research, under the National Plan for Complementary Investments to the NRRP, project “D34H—Digital Driven Diagnostics, prognostics and therapeutics for sustainable Health care” (project code: PNC0000001), Spoke 2: “Multilayer platform to support the generation of the Patients’ Digital Twin”, CUP: B53C22006170001, and Spoke 3: “Wearable technologies, sensors and biomarkers for care through Digital Twin approaches”, CUP: B53C22006100001. Senior research grant from the Research Scientific Foundation-Flanders (FWO) - G0D0520N, I-PREDICT: Innovative Physiology-based pharmacokinetic model to pREdict Drug exposure In neonates undergoing Cooling Therapy. KU CELSA research project (Central Europe Leuven Strategic Alliance, CELSA/24/022). Senior Clinical Investigatorship of the Research Foundation – Flanders (FWO) (18E2H24N). The funders played no role in study design, data collection, analysis and interpretation of data, or the writing of this manuscript. The work of C.R., A.B., and M.D.D.B. is supported by the Center of Excellence for Research DEWS, Design Methodologies for Embedded controllers, Wireless interconnect and System-on-chip, University of L’Aquila, Italy. The work of M.P. is supported by the Italian National Program Ph.D. Program in Autonomous Systems (DAuSy). We gratefully acknowledge the University Hospitals Leuven, KU Leuven and the iSi Health-KU Leuven Institute of Physics-based Modeling for In Silico Health, for enabling access to the neonatal clinical datasets used in this work. We also thank the organizers of the International Modeling Challenge Digital Twin Builder for Health Incubator: Building the Next Generation Physiologically based Framework for Precision Dosing of Renally Cleared Medicines in Neonates, whose initiative provided the opportunity to form the interdisciplinary team that carried out this study. The work of C.R., A.B., and M.D.D.B. is supported by the Center of Excellence for Research DEWS, Design Methodologies for Embedded controllers, Wireless interconnect and System-on-chip, University of L’Aquila, Italy. The work of M.P. is supported by the Italian National Program Ph.D. Program in Autonomous Systems (DAuSy). We gratefully acknowledge the University Hospitals Leuven, KU Leuven and the iSi Health-KU Leuven Institute of Physics-based Modeling for In Silico Health, for enabling access to the neonatal clinical datasets used in this work. We also thank the organizers of the International Modeling Challenge Digital Twin Builder for Health Incubator: Building the Next Generation Physiologically based Framework for Precision Dosing of Renally Cleared Medicines in Neonates, whose initiative provided the opportunity to form the interdisciplinary team that carried out this study.

Author contributions

M.P. and C.R. designed the study, implemented the computational model, performed all simulations and analyses, and prepared the manuscript. P.A. and K.A. collected and analyzed data, providing critical feedback ensuring the biological and clinical plausibility of the digital twin framework. M.D.D.B., N.A., and A.B. provided critical feedback to the methodology and contributed to technical, stylistic and implementation aspects. A.S. supervised the work and reviewed the manuscript. V.B. supervised the work, reviewed the manuscript, and obtained the resources. All authors discussed the results and implications and commented on the manuscript at all stages. All authors read and approved the final manuscript.

Data availability

The data that support the findings of this study are available from the Ethics Committee Research (EC Research) UZ/KU Leuven, but restrictions apply to the availability of these data, which were used under licence for the current study, and so are not publicly available. Data are however available from the authors upon reasonable request and with permission of EC Research UZ/KU Leuven. The underlying code for this study will be made accessible upon publication. The mechanistic ODEs specifying the PBPK-PD model were developed using the SimBiology toolbox v23.2, and the therapy control framework was developed using the Optimization toolbox v23.2 in MATLAB 2023b (Natick, MA, USA).

Code availability

The underlying code for this study is available at https://github.com/Michi354/EDT. The mechanistic ODEs specifying the PBPK-PD model were developed using the SimBiology toolbox v23.2, and the therapy control framework was developed using the Optimization toolbox v23.2 in MATLAB^® 2023b (Natick, MA, USA).

Competing interests

The authors declare no competing interests.

Footnotes

Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.

These authors contributed equally: Michela Prunella, Chiara Romano.

These authors jointly supervised this work: Anne Smits, Vitoantonio Bevilacqua.

Contributor Information

Chiara Romano, Email: chiara.romano1@graduate.univaq.it.

Alessandro Borri, Email: alessandro.borri@iasi.cnr.it.

Supplementary information

The online version contains supplementary material available at 10.1038/s41746-026-02558-w.

References

1.Wynn, J. L. & Wong, H. R. Pathophysiology of neonatal sepsis. In Fetal and Neonatal Physiology, 1536–1552.e10 https://linkinghub.elsevier.com/retrieve/pii/B9780323352147001529 (Elsevier, 2017).
2.World Health Organization. Global Report on the Epidemiology and Burden of Sepsis: Current Evidence, Identifying Gaps and Future Directions. Tech. Rep. (World Health Organization, 2020).
3.Natale, F., Bizzarri, B., Cardi, V. & De Curtis, M. Early and late onset sepsis in late preterm infants. Ital. J. Pediatr.40, A23, 1824–7288–40–S2–A23 (2014). [Google Scholar]
4.Andrews, J. M. Determination of minimum inhibitory concentrations. Nat. Rev. Microbiol.48, 5–16 (2001). [DOI] [PubMed] [Google Scholar]
5.Schmatz, M. et al. Surviving sepsis in a referral neonatal intensive care unit: association between time to antibiotic administration and in-hospital outcomes. J. Pediatr.217, 59–65.e1 (2020). [DOI] [PubMed] [Google Scholar]
6.Ngougni Pokem, P. et al. Dose optimization of β-lactam antibiotics in children: from population pharmacokinetics to individualized therapy. Expert Opin. Drug Metab. Toxicol.20, 787–804 (2024). [DOI] [PubMed] [Google Scholar]
7.Andersson, D. I. & Hughes, D. Microbiological effects of sublethal levels of antibiotics. Nat. Rev. Microbiol.12, 465–478 (2014). [DOI] [PubMed] [Google Scholar]
8.Mulinge, M. M., Mwanza, S. S., Kabahweza, H. M., Wamalwa, D. C. & Nduati, R. W. The impact of neonatal intensive care unit antibiotics on gut bacterial microbiota of preterm infants: a systematic review. Front. Microbiomes2, 1180565 (2023). [DOI] [PMC free article] [PubMed] [Google Scholar]
9.Ericson, J. E., Agthe, A. G. & Weitkamp, J.-H. Late-onset sepsis. Clin. Perinatol.52, 33–45 (2025). [DOI] [PubMed] [Google Scholar]
10.Flannery, D. D., Ramachandran, V. & Schrag, S. J. Neonatal early-onset sepsis. Clin. Perinatol.52, 15–31 (2025). [DOI] [PubMed] [Google Scholar]
11.McGann, C., Phyu, R., Bittinger, K. & Mukhopadhyay, S. Role of the microbiome in neonatal infection. Clin. Perinatol.52, 147–166 (2025). [DOI] [PubMed] [Google Scholar]
12.U.S. Food and Drug Administration. General Clinical Pharmacology Considerations for Neonatal Studies for Drugs and Biological Products (2022). Guidance for Industry (U.S. Food and Drug Administration, 2022).
13.Allegaert, K. & Van Den Anker, J. Neonatal drug therapy: the first frontier of therapeutics for children. Clin. Pharma Therapeutics98, 288–297 (2015). [DOI] [PMC free article] [PubMed] [Google Scholar]
14.Nathwani, D. et al. Value of hospital antimicrobial stewardship programs [ASPs]: a systematic review. Antimicrob. Resist Infect. Control8, 35 (2019). [DOI] [PMC free article] [PubMed] [Google Scholar]
15.Euteneuer, J. C., Kamatkar, S., Fukuda, T., Vinks, A. A. & Akinbi, H. T. Suggestions for model-informed precision dosing to optimize neonatal drug therapy. J. Clin. Pharma59, 168–176 (2019). [DOI] [PMC free article] [PubMed] [Google Scholar]
16.Rallis, D., Giapros, V., Serbis, A., Kosmeri, C. & Baltogianni, M. Fighting antimicrobial resistance in neonatal intensive care units: rational use of antibiotics in neonatal sepsis. Antibiotics12, 508 (2023). [DOI] [PMC free article] [PubMed] [Google Scholar]
17.Matcha, S. et al. Precision dosing of amikacin in term neonates using pharmacometric approach. Pediatr. Res98, 936–941 (2025). [DOI] [PMC free article] [PubMed] [Google Scholar]
18.Krzyzanski, W., Smits, A., Van Den Anker, J. & Allegaert, K. Population model of serum creatinine as time-dependent covariate in neonates. AAPS J.23, 86 (2021). [DOI] [PubMed] [Google Scholar]
19.Transtrum, M. K. & Qiu, P. Bridging mechanistic and phenomenological models of complex biological systems. PLoS Comput Biol.12, e1004915 (2016). [DOI] [PMC free article] [PubMed] [Google Scholar]
20.Feng, J., Lee, J., Vesoulis, Z. A. & Li, F. Predicting mortality risk for preterm infants using deep learning models with time-series vital sign data. npj Digit. Med.4, 108 (2021). [DOI] [PMC free article] [PubMed] [Google Scholar]
21.Ramirez, M. & Tolmasky, M. Amikacin: uses, resistance, and prospects for inhibition. Molecules22, 2267 (2017). [DOI] [PMC free article] [PubMed] [Google Scholar]
22.Wilbaux, M. et al. Pharmacometric approaches to personalize use of primarily renally eliminated antibiotics in preterm and term neonates. J. Clin. Pharma56, 909–935 (2016). [DOI] [PubMed] [Google Scholar]
23.Landersdorfer, C. B. & Nation, R. L. Limitations of antibiotic MIC-based PK-PD metrics: looking back to move forward. Front. Pharmacol.12, 770518 (2021). [DOI] [PMC free article] [PubMed] [Google Scholar]
24.Van Der Veer, M. A. A. et al. Amikacin dosing in neonates: evaluation of target attainment using a simplified and complex pharmacokinetic model-derived dosing regimen in clinical practice. Antimicrob. Agents Chemother.69, e01118–24 (2025). [DOI] [PMC free article] [PubMed] [Google Scholar]
25.Cristea, S. et al. Amikacin pharmacokinetics to optimize dosing in neonates with perinatal asphyxia treated with hypothermia. Antimicrob. Agents Chemother.61, e01282–17 (2017). [DOI] [PMC free article] [PubMed] [Google Scholar]
26.Smits, A., Annaert, P., Van Cruchten, S. & Allegaert, K. A physiology-based pharmacokinetic framework to support drug development and dose precision during therapeutic hypothermia in neonates. Front. Pharmacol.11, 587 (2020). [DOI] [PMC free article] [PubMed] [Google Scholar]
27.Gomez, L. et al. Implementation of a vancomycin dose-optimization protocol in neonates: impact on vancomycin exposure, biological parameters, and clinical outcomes. Antimicrob. Agents Chemother.66, e02191–21 (2022). [DOI] [PMC free article] [PubMed] [Google Scholar]
28.Colin, P. J., Jonckheere, S. & Struys, M. M. R. F. Target-controlled continuous infusion for antibiotic dosing: proof-of-principle in an in-silico vancomycin trial in intensive care unit patients. Clin. Pharmacokinet.57, 1435–1447 (2018). [DOI] [PMC free article] [PubMed] [Google Scholar]
29.Zwart, T. C. et al. Evaluation of amoxicillin and benzylpenicillin therapy in early-onset neonatal sepsis: a pharmacometric external validation and simulation study. J. Antimicrobial Chemother.80, 2214–2225 (2025). [DOI] [PMC free article] [PubMed] [Google Scholar]
30.D Denge, A., Vanga, Y. & Singh, R. Precision timing of beta-lactam administration in very low birth weight infants with suspected sepsis: a multi-national cohort study. J. Neonatal Surg.14, 919–931 (2025). [Google Scholar]
31.Boidin, C. et al. Amikacin initial dose in critically ill patients: a nonparametric approach to optimize a priori pharmacokinetic/pharmacodynamic target attainments in individual patients. Antimicrob. Agents Chemother.63, e00993–19 (2019). [DOI] [PMC free article] [PubMed] [Google Scholar]
32.Murphy, H. J. et al. Nephrotoxic medications and acute kidney injury risk factors in the neonatal intensive care unit: clinical challenges for neonatologists and nephrologists. Pediatr. Nephrol.35, 2077–2088 (2020). [DOI] [PubMed] [Google Scholar]
33.Egorov, A. M., Ulyashova, M. M. & Rubtsova, M. Y. Bacterial enzymes and antibiotic resistance. Acta Nat.10, 33–48 (2018). [PMC free article] [PubMed] [Google Scholar]
34.Yu, F., Wang, D., Zhang, H., Wang, Z. & Liu, Y. Evolutionary trajectory of bacterial resistance to antibiotics and antimicrobial peptides in Escherichia coli. mSystems10, e01700–24 (2025). [DOI] [PMC free article] [PubMed] [Google Scholar]
35.Ali, A., Imran, M., Sial, S. & Khan, A. Effective antibiotic dosing in the presence of resistant strains. PLOS ONE17, e0275762 (2022). [DOI] [PMC free article] [PubMed] [Google Scholar]
36.Romano, C., Benedetto, M. D. D. & Borri, A. Evolution-informed modeling and control of tumor growth. IEEE Trans. Automat. Sci. Eng. 1–1 https://ieeexplore.ieee.org/document/11146807/ (2025).
37.Charlebois, D. A. Quantitative systems-based prediction of antimicrobial resistance evolution. npj Syst. Biol. Appl.9, 40 (2023). [DOI] [PMC free article] [PubMed] [Google Scholar]
38.Camacho, E. F. & Bordons, C. Model predictive control. Advanced Textbooks in Control and Signal Processing 10.1007/978-0-85729-398-5 (Springer, 2007).
39.Prunella, M. et al. Pharmacometric and Digital Twin modeling for adaptive scheduling of combination therapy in advanced gastric cancer. Comput. Methods Prog. Biomed.270, 108919 (2025). [DOI] [PubMed] [Google Scholar]
40.Romano, C., Di Benedetto, M. D. & Borri, A. Stackelberg evolutionary games with modulated leadership: a three-agent framework for tumor-immune dynamics. IEEE Control Syst. Lett.9, 468–473 (2025). [Google Scholar]
41.Mohamed, A. F., Nielsen, E. I., Cars, O. & Friberg, L. E. Pharmacokinetic-pharmacodynamic model for gentamicin and its adaptive resistance with predictions of dosing schedules in newborn infants. Antimicrob. Agents Chemother.56, 179–188 (2012). [DOI] [PMC free article] [PubMed] [Google Scholar]
42.Melnyk, A. H., Wong, A. & Kassen, R. The fitness costs of antibiotic resistance mutations. Evolut. Appl.8, 273–283 (2015). [DOI] [PMC free article] [PubMed] [Google Scholar]
43.Khan, D. D. et al. A mechanism-based pharmacokinetic/pharmacodynamic model allows prediction of antibiotic killing from mic values for wt and mutants. J. Antimicrob. Chemother.70, 3051–3060 (2015). [DOI] [PubMed] [Google Scholar]
44.Kearns, G. L. et al. Developmental pharmacology – drug disposition, action, and therapy in infants and children. N. Engl. J. Med.349, 1157–1167 (2003). [DOI] [PubMed] [Google Scholar]
45.Rhodin, M. M. et al. Human renal function maturation: a quantitative description using weight and postmenstrual age. Pediatr. Nephrol.24, 67–76 (2009). [DOI] [PubMed] [Google Scholar]
46.Staub, E., Bolisetty, S., Allegaert, K. & Raaijmakers, A. Neonatal kidney function, injury and drug dosing: a contemporary review. Children12, https://api.semanticscholar.org/CorpusID:276875233 (2025). [DOI] [PMC free article] [PubMed]
47.Kiyoshige, A. et al. Association of neonatal serum creatinine concentration with maternal serum creatinine concentration and birth weight. Clin. Lab.69, http://www.clin-lab-publications.com/article/4404 (2023). [DOI] [PubMed]
48.Go, H. et al. Neonatal and maternal serum creatinine levels during the early postnatal period in preterm and term infants. PLOS ONE13, e0196721 (2018). [DOI] [PMC free article] [PubMed] [Google Scholar]
49.Miall, L. S. et al. Plasma creatinine rises dramatically in the first 48 hours of life in preterm infants. Pediatrics104, e76–e76 (1999). [DOI] [PubMed] [Google Scholar]
50.Van Wincoop, M., De Bijl-Marcus, K., Lilien, M., Van Den Hoogen, A. & Groenendaal, F. Effect of therapeutic hypothermia on renal and myocardial function in asphyxiated (near) term neonates: a systematic review and meta-analysis. PLOS ONE16, e0247403 (2021). [DOI] [PMC free article] [PubMed] [Google Scholar]
51.Krzyzanski, W. et al. A population model of time-dependent changes in serum creatinine in (near)term neonates with hypoxic-ischemic encephalopathy during and after therapeutic hypothermia. AAPS J.26, 4 (2023). [DOI] [PMC free article] [PubMed] [Google Scholar]
52.Hochreiter, S. & Schmidhuber, J. Long short-term memory. Neural Comput.9, 1735–1780 (1997). [DOI] [PubMed] [Google Scholar]
53.Glass, H. C. et al. Outcomes for extremely premature infants. Anesth. Analg.120, 1337–1351 (2015). [DOI] [PMC free article] [PubMed] [Google Scholar]
54.Zappitelli, M. et al. Developing a neonatal acute kidney injury research definition: a report from the NIDDK neonatal AKI workshop. Pediatr. Res.82, 569–573 (2017). [DOI] [PMC free article] [PubMed] [Google Scholar]
55.Allegaert, K. et al. Progression of the estimated glomerular filtration rate in asphyxiated neonates undergoing therapeutic hypothermia during the first 10 days of life. Pediatr. Nephrol.10.1007/s00467-025-06957-1 (2025). [DOI] [PMC free article] [PubMed]
56.Van Donge, T., Allegaert, K., Pfister, M., Smits, A. & Van Den Anker, J. Creatinine trends to detect ibuprofen-related maturational adverse drug events in neonatal life: a simulation study for the ELBW newborn. Front. Pharmacol.11, 610294 (2021). [DOI] [PMC free article] [PubMed] [Google Scholar]
57.Wildes, D. M. et al. A systematic review of the incidence of acute kidney injury in infants with patent ductus arteriosus. Curr. Pediatr. Rep.13, 4 (2025). [Google Scholar]
58.Hsu, K.-H. et al. Hemodynamic reference for neonates of different age and weight: a pilot study with electrical cardiometry. J. Perinatol.36, 481–485 (2016). [DOI] [PubMed] [Google Scholar]
59.Elstad, M., Liu, X. & Thoresen, M. Heart rate response to therapeutic hypothermia in infants with hypoxic-ischaemic encephalopathy. Resuscitation106, 53–57 (2016). [DOI] [PubMed] [Google Scholar]
60.Haycock, G. B., Schwartz, G. J. & Wisotsky, D. H. Geometric method for measuring body surface area: a height-weight formula validated in infants, children, and adults. J. Pediatrics93, 62–66 (1978). [DOI] [PubMed] [Google Scholar]
61.Witzany, C., Rolff, J., Regoes, R. R. & Igler, C. The pharmacokinetic-pharmacodynamic modelling framework as a tool to predict drug resistance evolution: this article is part of the Microbial Evolution collection. Microbiology169, 10.1099/mic.0.001368 (2023). [DOI] [PMC free article] [PubMed]
62.Hansen, E., Karslake, J., Woods, R. J., Read, A. F. & Wood, K. B. Antibiotics can be used to contain drug-resistant bacteria by maintaining sufficiently large sensitive populations. PLOS Biol.18, e3000713 (2020). [DOI] [PMC free article] [PubMed] [Google Scholar]
63.Butner, J. D. et al. Mathematical modeling of cancer immunotherapy for personalized clinical translation. Nat. Comput Sci.2, 785–796 (2022). [DOI] [PMC free article] [PubMed] [Google Scholar]
64.Bhattarai, S. K. et al. Commensal antimicrobial resistance mediates microbiome resilience to antibiotic disruption. Sci. Transl. Med.16, eadi9711 (2024). [DOI] [PMC free article] [PubMed] [Google Scholar]
65.Brepoels, P. et al. Antibiotic cycling affects resistance evolution independently of collateral sensitivity. Mol. Biol. Evol.39, msac257 (2022). [DOI] [PMC free article] [PubMed] [Google Scholar]
66.Claassen, K. et al. Development of a physiologically-based pharmacokinetic model for preterm neonates: evaluation with in vivo data. Curr. Pharmaceutical Design, https://www.semanticscholar.org/paper/Development-of-a-Physiologically-Based-Model-for-In-Claassen-Thelen/fe0b371b24ad9173b6f991dd560f9461a943f7d0 (2015). [DOI] [PubMed]
67.Jetton, J. G. et al. Incidence and outcomes of neonatal acute kidney injury (AWAKEN): a multicentre, multinational, observational cohort study. Lancet Child Adolesc. Health1, 184–194 (2017). [DOI] [PMC free article] [PubMed] [Google Scholar]
68.Zazo, H. et al. Physiologically-based pharmacokinetic modelling and dosing evaluation of gentamicin in neonates using PhysPK. Front. Pharmacol.13, 977372 (2022). [DOI] [PMC free article] [PubMed] [Google Scholar]
69.Anderson, B. & Holford, N. Mechanism-based concepts of size and maturity in pharmacokinetics. Annu. Rev. Pharmacol. Toxicol.48, 303–332 (2008). [DOI] [PubMed] [Google Scholar]
70.Fikac, L. Neonatal blood loss risks. Crit. Care Nurs. Q.42, 202–204 (2019). [DOI] [PubMed] [Google Scholar]
71.Mitropoulos, G., Scurry, J. & Cussen, L. Organ weight/bodyweight ratios: Growth rates of fetal organs in the latter half of pregnancy with a simple method for calculating mean organ weights. J. Paediatr. Child Health28, 236–239 (1992). [DOI] [PubMed] [Google Scholar]
72.Hughes, K. M. et al. Comparison of amikacin pharmacokinetics in neonates following implementation of a new dosage protocol. J. Pediatr. Pharmacol. Therapeutics22, 33–40 (2017). [DOI] [PMC free article] [PubMed] [Google Scholar]
73.Goodlet, K. J., Benhalima, F. Z. & Nailor, M. D. A systematic review of single-dose aminoglycoside therapy for urinary tract infection: Is it time to resurrect an old strategy? Antimicrob. Agents Chemother. 63, e02165-18 (2018). [DOI] [PMC free article] [PubMed]
74.Schwartz, G. J. et al. New equations to estimate GFR in children with CKD. J. Am. Soc. Nephrol.20, 629–637 (2009). [DOI] [PMC free article] [PubMed] [Google Scholar]

Associated Data

This section collects any data citations, data availability statements, or supplementary materials included in this article.

Supplementary Materials

Supplementary Information^{(2.7MB, pdf)}

Data Availability Statement

[CR1] 1.Wynn, J. L. & Wong, H. R. Pathophysiology of neonatal sepsis. In Fetal and Neonatal Physiology, 1536–1552.e10 https://linkinghub.elsevier.com/retrieve/pii/B9780323352147001529 (Elsevier, 2017).

[CR2] 2.World Health Organization. Global Report on the Epidemiology and Burden of Sepsis: Current Evidence, Identifying Gaps and Future Directions. Tech. Rep. (World Health Organization, 2020).

[CR3] 3.Natale, F., Bizzarri, B., Cardi, V. & De Curtis, M. Early and late onset sepsis in late preterm infants. Ital. J. Pediatr.40, A23, 1824–7288–40–S2–A23 (2014). [Google Scholar]

[CR4] 4.Andrews, J. M. Determination of minimum inhibitory concentrations. Nat. Rev. Microbiol.48, 5–16 (2001). [DOI] [PubMed] [Google Scholar]

[CR5] 5.Schmatz, M. et al. Surviving sepsis in a referral neonatal intensive care unit: association between time to antibiotic administration and in-hospital outcomes. J. Pediatr.217, 59–65.e1 (2020). [DOI] [PubMed] [Google Scholar]

[CR6] 6.Ngougni Pokem, P. et al. Dose optimization of β-lactam antibiotics in children: from population pharmacokinetics to individualized therapy. Expert Opin. Drug Metab. Toxicol.20, 787–804 (2024). [DOI] [PubMed] [Google Scholar]

[CR7] 7.Andersson, D. I. & Hughes, D. Microbiological effects of sublethal levels of antibiotics. Nat. Rev. Microbiol.12, 465–478 (2014). [DOI] [PubMed] [Google Scholar]

[CR8] 8.Mulinge, M. M., Mwanza, S. S., Kabahweza, H. M., Wamalwa, D. C. & Nduati, R. W. The impact of neonatal intensive care unit antibiotics on gut bacterial microbiota of preterm infants: a systematic review. Front. Microbiomes2, 1180565 (2023). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR9] 9.Ericson, J. E., Agthe, A. G. & Weitkamp, J.-H. Late-onset sepsis. Clin. Perinatol.52, 33–45 (2025). [DOI] [PubMed] [Google Scholar]

[CR10] 10.Flannery, D. D., Ramachandran, V. & Schrag, S. J. Neonatal early-onset sepsis. Clin. Perinatol.52, 15–31 (2025). [DOI] [PubMed] [Google Scholar]

[CR11] 11.McGann, C., Phyu, R., Bittinger, K. & Mukhopadhyay, S. Role of the microbiome in neonatal infection. Clin. Perinatol.52, 147–166 (2025). [DOI] [PubMed] [Google Scholar]

[CR12] 12.U.S. Food and Drug Administration. General Clinical Pharmacology Considerations for Neonatal Studies for Drugs and Biological Products (2022). Guidance for Industry (U.S. Food and Drug Administration, 2022).

[CR13] 13.Allegaert, K. & Van Den Anker, J. Neonatal drug therapy: the first frontier of therapeutics for children. Clin. Pharma Therapeutics98, 288–297 (2015). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR14] 14.Nathwani, D. et al. Value of hospital antimicrobial stewardship programs [ASPs]: a systematic review. Antimicrob. Resist Infect. Control8, 35 (2019). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR15] 15.Euteneuer, J. C., Kamatkar, S., Fukuda, T., Vinks, A. A. & Akinbi, H. T. Suggestions for model-informed precision dosing to optimize neonatal drug therapy. J. Clin. Pharma59, 168–176 (2019). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR16] 16.Rallis, D., Giapros, V., Serbis, A., Kosmeri, C. & Baltogianni, M. Fighting antimicrobial resistance in neonatal intensive care units: rational use of antibiotics in neonatal sepsis. Antibiotics12, 508 (2023). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR17] 17.Matcha, S. et al. Precision dosing of amikacin in term neonates using pharmacometric approach. Pediatr. Res98, 936–941 (2025). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR18] 18.Krzyzanski, W., Smits, A., Van Den Anker, J. & Allegaert, K. Population model of serum creatinine as time-dependent covariate in neonates. AAPS J.23, 86 (2021). [DOI] [PubMed] [Google Scholar]

[CR19] 19.Transtrum, M. K. & Qiu, P. Bridging mechanistic and phenomenological models of complex biological systems. PLoS Comput Biol.12, e1004915 (2016). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR20] 20.Feng, J., Lee, J., Vesoulis, Z. A. & Li, F. Predicting mortality risk for preterm infants using deep learning models with time-series vital sign data. npj Digit. Med.4, 108 (2021). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR21] 21.Ramirez, M. & Tolmasky, M. Amikacin: uses, resistance, and prospects for inhibition. Molecules22, 2267 (2017). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR22] 22.Wilbaux, M. et al. Pharmacometric approaches to personalize use of primarily renally eliminated antibiotics in preterm and term neonates. J. Clin. Pharma56, 909–935 (2016). [DOI] [PubMed] [Google Scholar]

[CR23] 23.Landersdorfer, C. B. & Nation, R. L. Limitations of antibiotic MIC-based PK-PD metrics: looking back to move forward. Front. Pharmacol.12, 770518 (2021). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR24] 24.Van Der Veer, M. A. A. et al. Amikacin dosing in neonates: evaluation of target attainment using a simplified and complex pharmacokinetic model-derived dosing regimen in clinical practice. Antimicrob. Agents Chemother.69, e01118–24 (2025). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR25] 25.Cristea, S. et al. Amikacin pharmacokinetics to optimize dosing in neonates with perinatal asphyxia treated with hypothermia. Antimicrob. Agents Chemother.61, e01282–17 (2017). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR26] 26.Smits, A., Annaert, P., Van Cruchten, S. & Allegaert, K. A physiology-based pharmacokinetic framework to support drug development and dose precision during therapeutic hypothermia in neonates. Front. Pharmacol.11, 587 (2020). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR27] 27.Gomez, L. et al. Implementation of a vancomycin dose-optimization protocol in neonates: impact on vancomycin exposure, biological parameters, and clinical outcomes. Antimicrob. Agents Chemother.66, e02191–21 (2022). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR28] 28.Colin, P. J., Jonckheere, S. & Struys, M. M. R. F. Target-controlled continuous infusion for antibiotic dosing: proof-of-principle in an in-silico vancomycin trial in intensive care unit patients. Clin. Pharmacokinet.57, 1435–1447 (2018). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR29] 29.Zwart, T. C. et al. Evaluation of amoxicillin and benzylpenicillin therapy in early-onset neonatal sepsis: a pharmacometric external validation and simulation study. J. Antimicrobial Chemother.80, 2214–2225 (2025). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR30] 30.D Denge, A., Vanga, Y. & Singh, R. Precision timing of beta-lactam administration in very low birth weight infants with suspected sepsis: a multi-national cohort study. J. Neonatal Surg.14, 919–931 (2025). [Google Scholar]

[CR31] 31.Boidin, C. et al. Amikacin initial dose in critically ill patients: a nonparametric approach to optimize a priori pharmacokinetic/pharmacodynamic target attainments in individual patients. Antimicrob. Agents Chemother.63, e00993–19 (2019). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR32] 32.Murphy, H. J. et al. Nephrotoxic medications and acute kidney injury risk factors in the neonatal intensive care unit: clinical challenges for neonatologists and nephrologists. Pediatr. Nephrol.35, 2077–2088 (2020). [DOI] [PubMed] [Google Scholar]

[CR33] 33.Egorov, A. M., Ulyashova, M. M. & Rubtsova, M. Y. Bacterial enzymes and antibiotic resistance. Acta Nat.10, 33–48 (2018). [PMC free article] [PubMed] [Google Scholar]

[CR34] 34.Yu, F., Wang, D., Zhang, H., Wang, Z. & Liu, Y. Evolutionary trajectory of bacterial resistance to antibiotics and antimicrobial peptides in Escherichia coli. mSystems10, e01700–24 (2025). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR35] 35.Ali, A., Imran, M., Sial, S. & Khan, A. Effective antibiotic dosing in the presence of resistant strains. PLOS ONE17, e0275762 (2022). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR36] 36.Romano, C., Benedetto, M. D. D. & Borri, A. Evolution-informed modeling and control of tumor growth. IEEE Trans. Automat. Sci. Eng. 1–1 https://ieeexplore.ieee.org/document/11146807/ (2025).

[CR37] 37.Charlebois, D. A. Quantitative systems-based prediction of antimicrobial resistance evolution. npj Syst. Biol. Appl.9, 40 (2023). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR38] 38.Camacho, E. F. & Bordons, C. Model predictive control. Advanced Textbooks in Control and Signal Processing 10.1007/978-0-85729-398-5 (Springer, 2007).

[CR39] 39.Prunella, M. et al. Pharmacometric and Digital Twin modeling for adaptive scheduling of combination therapy in advanced gastric cancer. Comput. Methods Prog. Biomed.270, 108919 (2025). [DOI] [PubMed] [Google Scholar]

[CR40] 40.Romano, C., Di Benedetto, M. D. & Borri, A. Stackelberg evolutionary games with modulated leadership: a three-agent framework for tumor-immune dynamics. IEEE Control Syst. Lett.9, 468–473 (2025). [Google Scholar]

[CR41] 41.Mohamed, A. F., Nielsen, E. I., Cars, O. & Friberg, L. E. Pharmacokinetic-pharmacodynamic model for gentamicin and its adaptive resistance with predictions of dosing schedules in newborn infants. Antimicrob. Agents Chemother.56, 179–188 (2012). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR42] 42.Melnyk, A. H., Wong, A. & Kassen, R. The fitness costs of antibiotic resistance mutations. Evolut. Appl.8, 273–283 (2015). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR43] 43.Khan, D. D. et al. A mechanism-based pharmacokinetic/pharmacodynamic model allows prediction of antibiotic killing from mic values for wt and mutants. J. Antimicrob. Chemother.70, 3051–3060 (2015). [DOI] [PubMed] [Google Scholar]

[CR44] 44.Kearns, G. L. et al. Developmental pharmacology – drug disposition, action, and therapy in infants and children. N. Engl. J. Med.349, 1157–1167 (2003). [DOI] [PubMed] [Google Scholar]

[CR45] 45.Rhodin, M. M. et al. Human renal function maturation: a quantitative description using weight and postmenstrual age. Pediatr. Nephrol.24, 67–76 (2009). [DOI] [PubMed] [Google Scholar]

[CR46] 46.Staub, E., Bolisetty, S., Allegaert, K. & Raaijmakers, A. Neonatal kidney function, injury and drug dosing: a contemporary review. Children12, https://api.semanticscholar.org/CorpusID:276875233 (2025). [DOI] [PMC free article] [PubMed]

[CR47] 47.Kiyoshige, A. et al. Association of neonatal serum creatinine concentration with maternal serum creatinine concentration and birth weight. Clin. Lab.69, http://www.clin-lab-publications.com/article/4404 (2023). [DOI] [PubMed]

[CR48] 48.Go, H. et al. Neonatal and maternal serum creatinine levels during the early postnatal period in preterm and term infants. PLOS ONE13, e0196721 (2018). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR49] 49.Miall, L. S. et al. Plasma creatinine rises dramatically in the first 48 hours of life in preterm infants. Pediatrics104, e76–e76 (1999). [DOI] [PubMed] [Google Scholar]

[CR50] 50.Van Wincoop, M., De Bijl-Marcus, K., Lilien, M., Van Den Hoogen, A. & Groenendaal, F. Effect of therapeutic hypothermia on renal and myocardial function in asphyxiated (near) term neonates: a systematic review and meta-analysis. PLOS ONE16, e0247403 (2021). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR51] 51.Krzyzanski, W. et al. A population model of time-dependent changes in serum creatinine in (near)term neonates with hypoxic-ischemic encephalopathy during and after therapeutic hypothermia. AAPS J.26, 4 (2023). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR52] 52.Hochreiter, S. & Schmidhuber, J. Long short-term memory. Neural Comput.9, 1735–1780 (1997). [DOI] [PubMed] [Google Scholar]

[CR53] 53.Glass, H. C. et al. Outcomes for extremely premature infants. Anesth. Analg.120, 1337–1351 (2015). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR54] 54.Zappitelli, M. et al. Developing a neonatal acute kidney injury research definition: a report from the NIDDK neonatal AKI workshop. Pediatr. Res.82, 569–573 (2017). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR55] 55.Allegaert, K. et al. Progression of the estimated glomerular filtration rate in asphyxiated neonates undergoing therapeutic hypothermia during the first 10 days of life. Pediatr. Nephrol.10.1007/s00467-025-06957-1 (2025). [DOI] [PMC free article] [PubMed]

[CR56] 56.Van Donge, T., Allegaert, K., Pfister, M., Smits, A. & Van Den Anker, J. Creatinine trends to detect ibuprofen-related maturational adverse drug events in neonatal life: a simulation study for the ELBW newborn. Front. Pharmacol.11, 610294 (2021). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR57] 57.Wildes, D. M. et al. A systematic review of the incidence of acute kidney injury in infants with patent ductus arteriosus. Curr. Pediatr. Rep.13, 4 (2025). [Google Scholar]

[CR58] 58.Hsu, K.-H. et al. Hemodynamic reference for neonates of different age and weight: a pilot study with electrical cardiometry. J. Perinatol.36, 481–485 (2016). [DOI] [PubMed] [Google Scholar]

[CR59] 59.Elstad, M., Liu, X. & Thoresen, M. Heart rate response to therapeutic hypothermia in infants with hypoxic-ischaemic encephalopathy. Resuscitation106, 53–57 (2016). [DOI] [PubMed] [Google Scholar]

[CR60] 60.Haycock, G. B., Schwartz, G. J. & Wisotsky, D. H. Geometric method for measuring body surface area: a height-weight formula validated in infants, children, and adults. J. Pediatrics93, 62–66 (1978). [DOI] [PubMed] [Google Scholar]

[CR61] 61.Witzany, C., Rolff, J., Regoes, R. R. & Igler, C. The pharmacokinetic-pharmacodynamic modelling framework as a tool to predict drug resistance evolution: this article is part of the Microbial Evolution collection. Microbiology169, 10.1099/mic.0.001368 (2023). [DOI] [PMC free article] [PubMed]

[CR62] 62.Hansen, E., Karslake, J., Woods, R. J., Read, A. F. & Wood, K. B. Antibiotics can be used to contain drug-resistant bacteria by maintaining sufficiently large sensitive populations. PLOS Biol.18, e3000713 (2020). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR63] 63.Butner, J. D. et al. Mathematical modeling of cancer immunotherapy for personalized clinical translation. Nat. Comput Sci.2, 785–796 (2022). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR64] 64.Bhattarai, S. K. et al. Commensal antimicrobial resistance mediates microbiome resilience to antibiotic disruption. Sci. Transl. Med.16, eadi9711 (2024). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR65] 65.Brepoels, P. et al. Antibiotic cycling affects resistance evolution independently of collateral sensitivity. Mol. Biol. Evol.39, msac257 (2022). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR66] 66.Claassen, K. et al. Development of a physiologically-based pharmacokinetic model for preterm neonates: evaluation with in vivo data. Curr. Pharmaceutical Design, https://www.semanticscholar.org/paper/Development-of-a-Physiologically-Based-Model-for-In-Claassen-Thelen/fe0b371b24ad9173b6f991dd560f9461a943f7d0 (2015). [DOI] [PubMed]

[CR67] 67.Jetton, J. G. et al. Incidence and outcomes of neonatal acute kidney injury (AWAKEN): a multicentre, multinational, observational cohort study. Lancet Child Adolesc. Health1, 184–194 (2017). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR68] 68.Zazo, H. et al. Physiologically-based pharmacokinetic modelling and dosing evaluation of gentamicin in neonates using PhysPK. Front. Pharmacol.13, 977372 (2022). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR69] 69.Anderson, B. & Holford, N. Mechanism-based concepts of size and maturity in pharmacokinetics. Annu. Rev. Pharmacol. Toxicol.48, 303–332 (2008). [DOI] [PubMed] [Google Scholar]

[CR70] 70.Fikac, L. Neonatal blood loss risks. Crit. Care Nurs. Q.42, 202–204 (2019). [DOI] [PubMed] [Google Scholar]

[CR71] 71.Mitropoulos, G., Scurry, J. & Cussen, L. Organ weight/bodyweight ratios: Growth rates of fetal organs in the latter half of pregnancy with a simple method for calculating mean organ weights. J. Paediatr. Child Health28, 236–239 (1992). [DOI] [PubMed] [Google Scholar]

[CR72] 72.Hughes, K. M. et al. Comparison of amikacin pharmacokinetics in neonates following implementation of a new dosage protocol. J. Pediatr. Pharmacol. Therapeutics22, 33–40 (2017). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR73] 73.Goodlet, K. J., Benhalima, F. Z. & Nailor, M. D. A systematic review of single-dose aminoglycoside therapy for urinary tract infection: Is it time to resurrect an old strategy? Antimicrob. Agents Chemother. 63, e02165-18 (2018). [DOI] [PMC free article] [PubMed]

[CR74] 74.Schwartz, G. J. et al. New equations to estimate GFR in children with CKD. J. Am. Soc. Nephrol.20, 629–637 (2009). [DOI] [PMC free article] [PubMed] [Google Scholar]

PERMALINK

Evolutionary digital twin framework for optimal aminoglycoside dosing in neonates with suspected sepsis

Michela Prunella

Chiara Romano

Alessandro Borri

Nicola Altini

Maria Domenica Di Benedetto

Pieter Annaert

Karel Allegaert

Anne Smits

Vitoantonio Bevilacqua

Abstract

Introduction

Fig. 1. Evolution-aware DT framework.

Results

A novel eco-evolutionary PD model links antibiotic exposure to adaptive bacterial response

Table 1.

Mechanistically informed calibration bridges neonatal physiology and eco-evolutionary dynamics

Fig. 2. Cohort-level selection probability and individual-level fitness dynamics under antibiotic exposure.

Data-driven modeling captures neonatal serum creatinine trajectories across heterogeneous clinical cohorts

Table 7.

Fig. 3. Postnatal sCr trajectories and data-driven versus mechanistic prediction performance across neonatal cohorts.

Comparative validation reveals superior fit to observed postnatal trajectories

The digital-twin framework accurately recapitulates observed amikacin PK and blood culture positivity patterns

Recapitulation of observed amikacin pharmacokinetics

Fig. 4. Integrated assessment of PBPK-predicted amikacin concentrations and PD-predicted bacterial dynamics across blood culture confirmed infections.

Table 2.

Recapitulation of blood culture positivity patterns

The virtual cohort reproduces maturational and pathophysiological neonatal patterns

Realism of renal function across pathophysiological conditions

Fig. 5. Distribution of sCr and eGFR as a function of non-maturational and maturational factors.

Table 3.

Realism of heart rate dynamics across maturational and clinical conditions

Fig. 6. Prevalence of PDA and heart rate responses under TH and inotrope exposure in the virtual cohort.

Realism of cardiovascular indicators and transitional hemodynamics

Table 4.

Evolutionary digital-twin optimization enables personalized and effective neonatal dosing

Iterative single-cycle optimization ensures robust and adaptive PK control under safety constraints

Optimization achieves high PK/PD target attainment while maintaining safety limits

Fig. 7. Population-level optimization results across the virtual cohort.

Pharmacodynamic outcomes reveal emergent sensitive dominance and ecological patterns

Fig. 8. Impact of MPC refinement on cohort resistance patterns and bacterial burden.

MPC refinement mitigates late-cycle bacterial rebound and improves long-horizon control

Concluding insights on cohort-level ecological control and future directions

Fig. 9. Cycle-wise prevalence of resistant and sensitive bacterial dominance.

DT-optimized dosing achieves improved bacterial control with clinically plausible schedules in real-world cohorts

Table 5.

Fig. 10. PK/PD target attainment on real neonatal cohorts under optimized dosing.

Table 6.

Discussion

Methods

Data

PBPK-PD model structure

Fig. 11. Schematic representation and equations of our PBPK-PD model.

PD model calibration procedure

LSTM modeling of serum creatinine dynamics

Fig. 12. LSTM structure and equations.

Model performance metrics

Optimization and simulation

Model-predictive control formulation

Comparison metrics between real and optimized schedules

Supplementary information

Acknowledgements

Author contributions

Data availability

Code availability

Competing interests

Footnotes

Contributor Information

Supplementary information

References

Associated Data

Supplementary Materials

Data Availability Statement

ACTIONS

PERMALINK

RESOURCES

Similar articles

Cited by other articles

Links to NCBI Databases