Summary
Background
Accurate prediction of the optimal dose for β-lactam antibiotics in neonatal sepsis is challenging. We aimed to evaluate whether a reliable clinical decision support system (CDSS) based on machine learning (ML) can assist clinicians in making optimal dose selections.
Methods
Five β-lactam antibiotics (amoxicillin, ceftazidime, cefotaxime, meropenem and latamoxef), commonly used to treat neonatal sepsis, were selected. The CDSS was constructed by incorporating the drug, patient, dosage, pharmacodynamic, and microbiological factors. The CatBoost ML algorithm was used to build the CDSS. Real-world studies were used to evaluate the CDSS performance. Virtual trials were used to compare the CDSS-optimized doses with guideline-recommended doses.
Findings
For a specific drug, by entering the patient characteristics and pharmacodynamic (PD) target (50%/70%/100% fraction of time that the free drug concentration is above the minimal inhibitory concentration [fT > MIC]), the CDSS can determine whether the planned dosing regimen will achieve the PD target and suggest an optimal dose. The prediction accuracy of all five drugs was >80.0% in the real-world validation. Compared with the PopPK model, the overall accuracy, precision, recall, and F1-Score improved by 10.7%, 22.1%, 64.2%, and 43.1%, respectively. Using the CDSS-optimized doses, the average probability of target concentration attainment increased by 58.2% compared to the guideline-recommended doses.
Interpretation
An ML-based CDSS was successfully constructed to assist clinicians in selecting optimal β-lactam antibiotic doses.
Funding
This work was supported by the National Natural Science Foundation of China; Distinguished Young and Middle-aged Scholar of Shandong University; National Key Research and Development Program of China.
Keywords: Neonates, Individual treatment, Machine-learning, β-lactam antibiotics, Clinical decision support system
Research in context.
Evidence before this study
Various factors determine the optimal dose of β-lactam antibiotics in neonates: (a) Developmental/maturation factors; (b) Pathogen factors. (c) Pharmacodynamics target; (d) Drug type and dosage regimen. Traditional population-based methods, while quantifying these factors to some degree, have problems with prognostic precision and accuracy.
Added value of this study
A machine learning-based clinical decision support system (CDSS) was successfully constructed to assist clinicians in making optimal dose selections for β-lactam antibiotics. The prediction accuracy of all five drugs was >80.0% in the real-world validation.
Implications of all the available evidence
For a specific drug, by entering patient characteristics (current weight, birth weight, gestational age, postnatal age, postmenstrual age, serum creatinine) and the PD target (50%/70%/100% fT > MIC [fraction of time that the free concentration is above the MIC]), the CDSS can determine whether the planned dosing regimen will achieve the PD target and suggest an optimal dose.
Introduction
Neonatal sepsis is an invasive disease with high rates of morbidity and mortality.1 β-lactam antibiotics are among the most frequently prescribed antibiotics for the treatment of neonatal sepsis.2 However, accurately predicting of the optimal dose of these antibiotics is difficult.
Various factors determine the optimal dose of β-lactam antibiotics. First, the rapid physiological changes and specific pathophysiology in neonates result in extensive intra- and inter-subject variability in drug disposition and clinical responses, respectively. Second, infectious pathogens and their corresponding MICs are the basis for selecting an appropriate antibiotic therapy. However, it is sometimes difficult to obtain positive cultures and/or antimicrobial susceptibility test results for bacteria that cause sepsis in neonates. Thirdly, the pharmacodynamics (PD) target of time-dependent β-lactam antibiotics is the fraction of the time during which the free antibiotic concentration remains above the minimum inhibitory concentration of the targeted pathogens (%fT > MIC).3 However, the specific target PD value in neonates remains controversial. There is no generally accepted PD target; as in recent neonatal studies, the PD target ranges from 40% fT > MIC to 100% fT > 4–5 x MIC. Lastly, the type of β-lactam antibiotics and their dosage regimens selected for the treatment of neonatal sepsis vary between centers and regions.
Although traditional population-based methods quantify these factors to some degree, there are problems with their prognostic precision and accuracy.4 Therefore, a new advanced approach capable of integrating multiple factors is required to accurately prediction of individual optimal doses. In recent years, machine learning (ML) has become increasingly popular in clinical applications The mechanisms of ML are unique and sensitive to multiple factors and are now becoming the key to solving difficult problems.
Thus, this study aimed to evaluate whether a clinical decision support system (CDSS) can be constructed based on ML integration of multiple factors to assist clinicians in optimal dose selection, using five beta-lactam antibiotics (ceftazidime, cefotaxime, meropenem, latamoxef, and amoxicillin) commonly used in the treatment of neonatal sepsis, as “proof of concept”.
Methods
The methods of this study consisted of three steps: CDSS construction, external validation of the CDSS in real-world study, and CDSS-based dose optimization. A schematic of this study was shown in Fig. 1.
Fig. 1.
Diagram illustrating the construction, evaluation, optimization, and application of the clinical decision support system. CDSS: Clinical Decision Support System; PopPK: population pharmacokinetic; ML: machine learning; CL: clearance; V1: central volume of distribution; V2: peripheral volume of distribution; Q: inter-compartment clearance.
CDSS construction
The CDSS was achieved through the construction of ML datasets and subsequent ML analysis to develop predictive models.
ML dataset construction
Information on neonates receiving antibacterial treatments from previous neonatal studies was used to establish a virtual neonate dataset.5, 6, 7, 8, 9, 10, 11, 12, 13 The dataset included information on 2569 neonates. The mean (SD) values of current weight (CW) and postmenstrual age (PMA) were 2463 (1023) [range 450–9370] g and 36.9 (4.43) [range 28.4–48.0] weeks. Extremely preterm births (GA < 28 weeks) were excluded from the virtual dataset. Detailed patient information was shown in Table 1.
Table 1.
Statistical information for virtual neonates.
| n | Mean (SD) | Median (range) | |
|---|---|---|---|
| Patients characteristics | 2569 | ||
| GA (weeks) | 34.8 (4.34) | 35.3 (28.0–43.4) | |
| PNA (days) | 15.0 (13.6) | 11.0 (1.00–91.0) | |
| PMA (weeks) | 36.9 (4.43) | 37.4 (28.4–48.0) | |
| CW (g) | 2463 (1023) | 2400 (450–9370) | |
| BW (g) | 2345 (828) | 2056 (1000–4850) | |
| CREA (μmol/L) | 48.7 (28.6) | 45.0 (1.10–518) |
CW: current weight; BW: birth weight; GA: gestational age; PNA: postnatal age; PMA: postmenstrual age; CREA: serum creatinine.
For each of the five drugs, published population pharmacokinetic (PopPK) models were searched and selected that could cover the entire age range of the neonatal population. Based on the PopPK model, Monte Carlo simulations were performed for individual concentrations under different scenarios (see below) using the virtual neonatal dataset. In this study, considering the actual clinical medication scenarios, the following four aspects were considered in the setting of the simulation scenarios: (1) Drug categories. In this study, different categories of β-lactam antibiotics were selected for clinical use in neonates, including penicillins (amoxicillin), cephalosporins (ceftazidime and cefotaxime), carbapenems (meropenem) and oxacephalosporins (latamoxef). They are primarily eliminated by the kidneys via time-dependent bacterial killing mechanisms. In adults, the protein binding rates of ceftazidime, cefotaxime, meropenem, latamoxef, and amoxicillin were 10.0%, 40.0%, 2.0%, 52.0%, and 10.0%, respectively.14, 15, 16, 17, 18 Because of the lack of data on protein binding in neonates, the adult values were used. The free concentration of a drug that is effective in the body is equal to the total drug concentration multiplied by the proportion that is not bound to the protein. (2) Patient Characteristics. The large dataset of virtual neonates covered the entire age range of neonates and young infants. Covariates reflecting the growth (size), renal maturity, and renal function of the neonates, that is, birth weight (BW), current weight (CW), gestational age at birth (GA), postmenstrual age (PMA), postnatal age (PNA), serum creatinine concentration (CREA), and albumin (ALB), were included in the dataset. (3) Dosage regimen. The dose range selected for the simulation set covered the drug's routine dosing regimens for neonates in clinical practice. The simulated dose ranges for ceftazidime, cefotaxime, meropenem, latamoxef, and amoxicillin were 20–50, 20–60, 10–40, 15–40, and 20–50 mg/kg, respectively. During these simulations, the dose was increased from the lower limit of the range of 5 units to the upper limit. Two commonly used dosing frequencies in the clinical settings were selected: Q12h and Q8h. All five drugs were administered as a 30 min intravenous infusion. (4) Sampling time. The PD target values (%fT > MIC) of β-lactam antibiotics are still controversial. We selected three common PD targets, 50%, 70% and 100% fT > MIC. Therefore, we simulated the steady-state concentrations at 50%, 70%, and 100% dosing interval time points. The number of simulation scenarios was determined by multiplying the number of drug categories, patients, dosage regimens, and sampling times. For each virtual neonate in a simulation scenario, 1000 simulated concentrations were obtained from 1000 Monte Carlo simulations. The median concentration was then calculated and used as the final simulated concentration for each individual in the given scenario.
After obtaining the simulated concentrations under different simulation scenarios, it was essential to use the pathogen MIC values of the pathogen to determine whether the concentrations were adequate (i.e., whether the simulated concentrations covered the MIC of the pathogen). Therefore, the choice of the MIC was crucial. This study investigated relevant studies to identify the common pathogens causing early-onset sepsis (EOS) and late-onset sepsis (LOS). The MIC values of these five antibacterial agents against pathogenic bacteria were determined according to the European Committee on Antimicrobial Susceptibility Testing (EUCAST), Clinical and Laboratory Standards Institute (CLSI) and other studies.19, 20, 21, 22, 23, 24
A new binary variable, “TAR”, was formed by comparing the simulated concentration with the pathogen MIC value to determine whether the individual concentration under different simulation scenarios met the target. If the simulated concentration met the target, then “TAR” = 1; if it does not, then “TAR” = 0. For example, if a virtual neonate was administered a dosing regimen of amoxicillin 20 mg/kg Q12h, the steady-state simulated concentration at the 50% dosing interval was 7 mg/L. If the corresponding pathogen MIC was 8 mg/L, the target was not met in the current situation, and “TAR” = 0. “TAR” was used as a prediction target in building subsequent ML analysis.
ML analysis
The binary variable “TAR” was used as a prediction target. The following potential predictors were included in the simulation scenarios mentioned above, the chosen drug, patient characteristics (BW, CW, PNA, PMA, GA, CREA, and ALB), dosage and frequency (Q12h/Q8h) of the chosen drug, and sampling time (50/70/100% dosing interval). The simulation dataset was divided into training (80.0%) and the test (20.0%) datasets. Five-fold nested cross-validation (NeCV) was applied to tune the hyper-parameters and evaluate the model performance. NeCV has gained widespread acceptance as “state-of-the-art” in the ML field as it is a (nearly) objective model assessment approach for predicting the genuine error.25,26 The inner cross-validation selected the best hyper-parameters, and the performance of the model was assessed by the outer cross-validation, utilizing the best parameters from the inner cross-validation selection. Detailed information on the NeCV method used in this study was shown in Fig. 2.
Fig. 2.
Nested cross-validation method used in this study.
Categorical Boosting (CatBoost) was used for ML analysis.27 CatBoost is a fast, scalable, and high-performance algorithm for gradient boosting of decision trees.28 CatBoost implements oblivious decision trees, thereby restricting the features split per level to one, which helps reduce the prediction time. It effectively handled categorical features using ordered target statistics. The accuracy, precision, recall, and balanced score (F1-Score) were calculated to evaluate the predictive performance. The Shapley additive explanations (SHAP) were calculated and visualized to evaluate the contribution of each covariate.29 All the ML analyses were implemented in Anaconda using Python 3.7. The final ML prediction models for the five β-lactam antibiotics were packaged and integrated to form the CDSS.
External validation of CDSS in real-world study
Evaluation of CDSS predictive performance
Real-world data were obtained from five hospitals. Neonates receiving anti-infective treatment with the five previously mentioned β-lactams were included. Neonatal information was collected based on the parameters of the trained ML model, such as drug categories, patient characteristics, dosage regimen (dosage and frequency), sampling time, and individual concentrations.
The real target (i.e., “TAR”) of the real-world validation dataset was set based on whether the actual individual concentrations exceeded the MIC. It can be categorized into two situations. (1) Standard scenario: Individual steady-state concentrations were sampled at the target time point (50% or 70% or 100% dosing interval ± 0.5 h). It can directly determine whether the concentration at the target time point has achieved the target. (2) Inference scenario: If the sampling time point for individual steady-state concentration was before the target time point, and the concentration was less than the MIC value, then “TAR” = 0; however, if the sampling time point for individual concentrations was after the target time point, and the concentration was greater than the MIC value, then “TAR” = 1.
Using the CDSS, a prediction was made regarding whether individual concentrations would meet the target under the actual dosing regimen. The results were compared with a real target of a real-world validation dataset, and the prediction accuracy, precision, recall, and F1-Score were calculated.
Predictive performance of CDSS compared to the PopPK model
In this section, we compared the predictive performances of the CDSS and PopPK models. For the PopPK model (the PopPK model selected for each drug was the same as that in Section 2.1), individual concentrations were predicted based on typical population estimates of the model and all available patient covariates and dosing information. The prediction accuracy, precision, recall, and F1-Score were calculated as described previously, and the results were compared with those of the CDSS.
CDSS-based dose optimization
Virtual trials were used to compare the CDSS-optimized doses with the guideline-recommended doses. The CDSS-optimized dose was obtained by optimizing the individual dose using the CDSS. The sources of the guideline-recommended doses were as follows: (1) Manual of Childhood Infections: The Blue Book (BLUE)30; (2) British National Formulary for Children (BNFC)31; and (3) Food and Drug Administration label (FDAL).
Virtual trials were conducted using an external validation dataset. Monte Carlo simulations were conducted based on the PK parameters of the published PopPK model, and 1000 simulations were performed for the CDSS-optimized and guideline-recommended dose regimens. The individual steady-state concentrations were predicted for each simulated virtual neonate at 70% of the dosing interval. The probability of target concentration attainment (PTA) was calculated as the percentage of virtual neonates meeting the PD target (steady-state individual concentrations in the range of 1–8 × MIC, where 8 × MIC was used as the toxicity threshold32).
Ethic
All the data were obtained from previous studies. These studies were approved by the institutional ethics committees of different hospitals: Beijing Children’s Hospital (2020-k-163), Beijing Obstetrics and Gynecology Hospital (2020-KY-045-01), Shandong Province Qianfoshan Hospital (2017S021), Children’s Hospital of Hebei Province (No.127), and Affiliated Hospital of Xuzhou Medical University (XYFY2019-KL194-01). Parental informed consent to participate was obtained for each newborn.
Role of funders
The funders played no role in the study design, data collection, data analysis, data interpretation, and paper writing.
Results
CDSS construction
Five published PopPK models (corresponding to five drugs) covering the entire age range of neonates and young infants were selected for the simulation process.22,33, 34, 35, 36 The PopPK mathematical model is described in detail in Supplementary Material S1. The characteristics of the neonates enrolled for modeling and the mathematical models of the selected PopPK models for the five drugs were presented in Table 2 and Supplementary Material S1.
Table 2.
Patient characteristics extracted from the published studies.
| Ceftazidime | Cefotaxime | Meropenem | Latamoxef | Amoxicillin | |
|---|---|---|---|---|---|
| Patients | 146 | 100 | 188 | 128 | 187 |
| Samples | 203 | 185 | 780 | 165 | 224 |
| BW (kg) | 3.00 (0.740–4.65) | 1.42 (0.512–3.99) | 1.08 (0.330–4.77) | 3.10 (1.01–4.58) | 3.05 (1.04–4.60) |
| CW (kg) | 3.08 (0.900–4.50) | 1.65 (0.530–4.20) | / | 3.22 (1.00–4.60) | 3.21 (1.06–4.58) |
| GA (weeks) | 38.6 (26.0–43.4) | 31.5 (23.0–42.0) | 28.0 (23.0–40.0) | 38.3 (27.3–41.4) | 38.1 (28.3–41.4) |
| PNA (days) | 11.0 (1.00–81.0) | 9.0 (1.00–69.0) | 21.0 (1.00–92.0) | 8.00 (1.00–54.0) | 7.00 (1.00–37.0) |
| PMA (weeks) | 40.3 (26.1–47.4) | 33.0 (25.0–44.0) | 33.0 (24.0–51.0) | 39.7 (28.4–46.1) | 39.0 (28.4–46.3) |
| CREA (μmol/L) | 33.8 (8.0–201.3) | 44.0 (13.0–226.0) | / | / | 38.0 (8.80–156) |
| ALB (g/L) | 31.7 (18.4–42.0) | 27.0 (17.0–38.0) | 23 (/) | / | / |
Patient demographic characteristics and clearance values are presented as median (range).
CW: current weight; BW: birth weight; GA: gestational age; PNA: postnatal age; PMA: postmenstrual age; CREA: serum creatinine; ALB: albumin.
According to national and international references, Escherichia coli, Klebsiella pneumoniae, Group B Streptococcus, Staphylococcus aureus, and Coagulase-negative Staphylococcus are the most common pathogens that cause neonatal sepsis.37, 38, 39, 40, 41, 42, 43, 44, 45, 46, 47 The MIC values of the five drugs for different pathogens were presented in Table 3. Then, 8, 8, 8, 4, and 4 mg/L were selected as the final MICs for amoxicillin, ceftazidime, cefotaxime, meropenem, and latamoxef, respectively, to identify the most common pathogens. Thus, these selected MICs were used to determine whether the individual concentrations simulated under different scenarios met the PD targets.
Table 3.
The MIC values of different pathogens for the five drugs.
| MIC values of five drugs |
|||||
|---|---|---|---|---|---|
| AMX | CTD | COX | MERO | MOX | |
| Pathogens | |||||
| E.coli | 8 | <4/8 | 1 | <4 | 0.12 |
| KP | 8 | <4/8 | 1 | <4 | 1 |
| CoNS | -a | -a | 4 | 4 | 4 |
| STA | -a | -a | 4/8 | 4 | 4 |
| GBS | 0.25 | -a | 0.5 | 0.25 | -b |
E.coli: Escherichia coli; KP: Klebsiella pneumoniae; CoNS: Coagulase-negative Staphylococcus; STA: Staphylococcus aureus; GBS: Group B Staphylococcus. AMX: amoxicillin; CTD: ceftazidime; COX: cefotaxime; MERO: meropenem; MOX: latamoxef.
Coagulase-negative Staphylococcus, Staphylococcus aureus, and Group B Staphylococcus were usually resistant to ceftazidime, amoxicillin.
: Not found.
Using the CatBoost algorithm, the prediction accuracy, accuracy, recall, and F1-Score for all five drugs exceeded 98.0%. Machine learning analysis was applied separately to each of the five drugs, and the results were presented in Table 4. Different drugs have different covariate importance rankings. For ceftazidime, cefotaxime, and latamoxef, the dosage regimen (dose and frequency) and PD target were the three most important covariates, followed by neonatal maturation and size covariates. For meropenem, PD target, frequency and PMA were the most important three covariates. For amoxicillin, PMA, frequency and dose were the most important three covariates. CREA was almost tricky to help prediction, except for meropenem. The relative importance of the covariates influencing the predictions for each of the five drugs was presented in Fig. 3.
Table 4.
Machine learning analysis results (Test set and Real-world validation data results).
| AMX | CTD | COX | MERO | MOX | |
|---|---|---|---|---|---|
| Test set results | |||||
| ML | |||||
| Accuracy (%) | 98.8 | 98.9 | 99.0 | 98.7 | 98.7 |
| Precision (%) | 98.5 | 99.2 | 99.3 | 98.9 | 98.9 |
| Recall (%) | 98.2 | 99.3 | 99.4 | 99.0 | 99.1 |
| F1 score (%) | 98.3 | 99.3 | 99.3 | 99.0 | 99.0 |
| Real world validation data results | |||||
| ML (CDSS) | |||||
| Accuracy (%) | 96.2 | 84.1 | 90.4 | 83.1 | 88.5 |
| Precision (%) | 82.4 | 85.5 | 93.8 | 86.5 | 93.4 |
| Recall (%) | 82.4 | 87.7 | 95.8 | 86.5 | 89.5 |
| F1score (%) | 82.4 | 86.6 | 94.8 | 86.5 | 91.4 |
| PopPK | |||||
| Accuracy (%) | 87.9 | 82.6 | 77.9 | 73.5 | 78.4 |
| Precision (%) | 40.0 | 85.2 | 90.5 | 89.5 | 90.1 |
| Recall (%) | 23.5 | 85.2 | 80.7 | 65.4 | 76.8 |
| F1score (%) | 29.6 | 85.2 | 85.4 | 75.6 | 83.0 |
AMX: amoxicillin; CTD: ceftazidime; COX: cefotaxime, MERO: meropenem; MOX: latamoxef.
Fig. 3.
SHAP results of model covariates. Each point represents a sample. A higher intensity of the red color indicates a more relevant covariate whereas a higher intensity of the blue color indicates a less relevant covariate. If most SHAP values of the red points are located in the area greater than 0, and most SHAP values of the blue points are located in the area less than 0, it means that the covariate has a positive correlation with the target variable and vice versa. BW: birth weight; CW: current weight; GA: gestational age at birth; PMA: postmenstrual age; PNA: postnatal age; CREA: Serum creatinine concentration; ALB: albumin. (A) Ceftazidime; (B) cefotaxime; (C) meropenem; (D) latamoxef; (E) amoxicillin.
External validation of CDSS in real-world study
Predictive performance of CDSS
In total, 621 blood samples from 621 patients were included in the validation dataset. The necessary information was collected based on the covariates of the final ML model, including GA, PMA, PNA, BW, CW, CREA, AMT, FRE, FT, and TAR (with the additional collection of ALB for meropenem). Demographic information, drug treatment details, and essential information for model predictions were summarized in Table 5.
Table 5.
Information of the real-world validation dataset.
| Amoxicillin | Ceftazidime | Cefotaxime | Meropenem | Latamoxef | |
|---|---|---|---|---|---|
| Patient (samples) | 157 (157) | 138 (138) | 104 (104) | 83 (83) | 139 (139) |
| GA (weeks)a | 38.0 (28.1–41.3) | 37.9 (26.0–43.4) | 32.0 (23.0–42.0) | 35.3 (26.3–42.1) | 37 (27.3–41.1) |
| PMA (weeks)a | 38.9 (28.4–44.1) | 40.0 (26.1–47.4) | 34.2 (26.4–44.4) | 37.1 (25.6–48.0) | 38.3 (31.0–44.1) |
| PNA (days)a | 7.0 (1.0–37.0) | 13.0 (1.0–81.0) | 8.5 (1.0–108.0) | 23.0 (4.0–116.0) | 7.0 (1.0–54.0) |
| BW (kg)a | 3.0 (1.2–4.2) | 2.8 (0.7–4.7) | 2.7 (1.4–3.5) | 1.9 (0.5–3.9) | 2.7 (1.2–4.1) |
| CW (kg)a | 3.2 (1.2–4.3) | 2.9 (0.9–4.8) | 1.9 (0.5–4.2) | 2.1 (1.1–5.3) | 2.7 (1.1–4.2) |
| CREA (μmol/L)a | 36.0 (8.8–167.9) | 34.4 (8.0–201.0) | 45.0 (12.0–226.0) | 25.2 (1.3–126.0) | 41.0 (4.5–103.0) |
| ALB (g/L)a | / | / | / | 2.3 (1.5–4.5) | / |
| Treatment | |||||
| Dose (mg/dose)a | 75.0 (25.0–110.0) | 90.0 (30.0–140.0) | 100 (28.0–208.0) | 55.0 (18.0–150.0) | 85.0 (40.0–140.0) |
| Dose (mg/kg/dose)a | 23.7 (16.7–32.1) | 29.1 (21.2–80.0) | 49.5 (31.7–105.0) | 19.8 (8.80–85.7) | 30.0 (24.0–56.1) |
| FREb | Q12h: 63.7% Q8h: 36.3% |
Q12h: 60.1% Q8h: 39.9% |
Q12h: 52.9% Q8h: 47.1% |
Q12h: 28.9% Q8h: 71.1% |
Q12h: 69.8% Q8h: 30.2% |
| FTb | 1: 13.4% 2: 8.9% 3: 77.7% |
1: 58.7% 2: 3.6% 3: 37.7% |
1: 40.4% 2: 30.8% 3: 28.8% |
1: 47.0% 2: 19.3% 3: 33.7% |
1: 15.2% 2: 46.0% 3: 38.8% |
| TARb | 0: 73.2% 1: 26.8% |
0: 41.3% 1: 58.7% |
0: 20.2% 1: 79.8% |
0: 37.3% 1: 62.7% |
0: 31.7% 1: 68.3% |
CW: current weight; BW: birth weight; GA: gestational age; PNA: postnatal age; PMA: postmenstrual age; CREA: serum creatinine; ALB: albumin; FRE: Dosage Frequency; FT: PD targets, 50/70/100% fT > MIC, 1,2,3 represent 50%, 70% and 100% respectively; TAR: prediction target, concentration meets the standard as 1, does not meet the standard as 0.
The results were shown in median (range).
The results are presented as proportions for each category.
Amoxicillin was administered to 157 patients, with 73.2% of the patients having real-world concentrations below the target (i.e., TAR < 0). Among the patients, 63.7% (n = 100) were administered a Q12h dosing regimen with a median (range) dose of 23.7 (16.7–32.1) mg/kg. Ceftazidime was administered to 138 patients, with 41.3% of the patients having real-world concentrations below the target (i.e., TAR < 0). Among the patients, 60.1% (n = 100) were administered with a Q12h dosing regimen, with a median (range) dose of 29.1 (21.2–80.0) mg/kg. Cefotaxime was administered to 104 patients, with 20.2% of the patients having real-world concentrations below the target (i.e., TAR < 0). Among the patients, 52.9% (n = 100) were administered with a Q12h dosing regimen, with a median (range) dose of 49.5 (31.7–105.0) mg/kg. Meropenem was administered to 83 patients, with 37.3% of the patients having real-world concentrations below the target (i.e., TAR < 0). Among the patients, 28.9% (n = 100) were administered with a Q12h dosing regimen, with a median (range) dose of 19.8 (8.80–85.7) mg/kg. Latamoxef was administered to 139 patients, with 31.7% of the patients having real-world concentrations below the target (i.e., TAR < 0). Among these patients, 69.8% (n = 100) received a Q12h dosing regimen, with a median (range) dose of 30.0 (24.0–56.1) mg/kg.
For the CDSS, the accuracies were 96.2%, 84.1%, 90.4%, 83.1%, and 88.5% for amoxicillin ceftazidime, cefotaxime, meropenem, and latamoxef, respectively. The precisions were 82.4%, 93.8%, 86.5%, 93.4%, and 85.5%, respectively. The recalls were 82.4%, 95.8%, 86.5%, 89.5%, and 87.7%, respectively. The F1-Scores were 82.4%, 94.8%, 86.5%, 91.4%, and 86.6%, respectively.
Predictive performance of CDSS compared to the PopPK model
For the PopPK model, the accuracies were 87.9%, 82.6%, 77.9%, 73.5%, and 78.4% for amoxicillin ceftazidime, cefotaxime, meropenem, and latamoxef, respectively. The precisions were 40.0%, 85.2%, 90.5%, 89.5% and 90.0%, respectively. The recalls were 23.5%, 85.2%, 80.7%, 65.4% and 76.8%, respectively. The F1-Scores were 29.6%, 85.2%, 85.4%, 75.6% and 83.0%, respectively.
Using the CDSS, the accuracy improved by 9.4%, 1.8%, 16.1%, 13.1% and 12.9% for amoxicillin ceftazidime, cefotaxime, meropenem, and latamoxef, respectively, compared to the PopPK model. Overall accuracy improved by 10.7%. The overall prediction precision, recall, and F1-Score of the CDSS (ML model) improved by 22.1%, 64.2, and 43.1%, respectively. The detailed validation results were shown in Table 4.
CDSS-based dose optimization
The guidelines recommend doses of beta-lactam antibiotics for anti-infective therapy in neonates, as shown in Table 6 (excluding indications for neonatal meningitis). Using the CDSS-optimized dosing regimen, the PTA for amoxicillin, ceftazidime, cefotaxime, meropenem, and latamoxef increased by 113.5%, 113.5%, 25.3%, 30.7%, and 8.1%, respectively, relative to the guideline-recommended dosing regimen (mean PTA). The PTAs for the CDSS-optimized dose regimen and the guideline-recommended dose regimen in neonatal patients were shown in Table 7.
Table 6.
The guidelines recommend doses of beta-lactam antibiotics for anti-infective therapy in neonates.
| Amoxicillin | Ceftazidime | Cefotaxime | Meropenem | Latamoxef | |
|---|---|---|---|---|---|
| BLUE | PNA <7 day: 30–60 mg/kg Q12h PNA ≥ 7 day: 30–60 mg/kg Q8h |
PNA<7 day: 25 mg/kg Q12h PNA ∼ 7–21 day: 25 mg/kg Q8h PNA ∼21–28 day: 25 mg/kg Q6h |
PNA <7 day: 25–50 mg/kg Q24h PNA ∼ 7–21 day: 25–50 mg/kg Q12h PNA ≥ 21 day: 25–50 mg/kg Q8h |
PNA <7 day: 40 mg/kg Q12h PNA ≥ 7 day: 40 mg/kg Q8h |
/ |
| BNFC | PNA <7 day: 30–50 mg/kg Q12h PNA ≥ 7 day: 30–50 mg/kg Q8h |
PNA<7 day: 25 mg/kg Q12h PNA ∼ 7–21 day: 25 mg/kg Q8h PNA ∼21–28 day: 25 mg/kg Q6h |
PNA <7 day: 25 mg/kg Q24h PNA ∼ 7–21: 25 mg/kg Q12h PNA ∼21–28 day: 25 mg/kg Q8h |
PNA <7 day: 20 mg/kg Q12h PNA ≥ 7 day: 20 mg/kg Q8h |
/ |
| FDAL | 30 mg/kg/day Q12h | PNA <7 day: 50 mg/kg Q12h PNA ∼7–28 day: 50 mg/kg Q8h |
30 mg/kg Q12h | / | 40–80 mg/kg/day Q6–12 ha |
The relevant dose was not retrieved and was substituted using the Chinese drug instruction dose. BLUE: The Blue Book; BNFC: British National Formulary for Children; FDAL: Food and Drug Administration Label; PNA: postnatal age.
Table 7.
The probability of target concentration attainment of CDSS-optimized dose and guideline-recommended dose.
| Amoxicillin | Ceftazidime | Cefotaxime | Meropenem | Latamoxef | |
|---|---|---|---|---|---|
| CDSS (%) | 77.5 | 92.0 | 77.9 | 77.1 | 82.7 |
| BLUE (%) | 52.9 | 55.5 | 56.7 | / | / |
| BNFC (%) | 47.1 | 29.0 | 56.7 | 59.0 | / |
| FDAL (%) | 8.8 | 44.9 | 73.1 | / | 76.5 |
| MEANG (%) | 36.3 | 43.1 | 62.2 | 59.0 | 76.5 |
CDSS: clinical decision-support system; BLUE: The Blue Book; BNFC: British National Formulary for Children; FDAL: Food and Drug Administration Label; MEANG: Mean values for the guideline-recommended doses.
Software and website construction
The CDSS was converted into software and a website for application in clinical practice. The CDSS software and website were designed to operate in a mode that could automatically invoke updates, thereby enabling dose prediction without the need for multiple repeated attempts. In other words, when the CDSS predicts a current dose failure, it not only notifies the user of the inadequacy but also automatically searches for the optimal dose (by setting the dose range and single-step incremental doses). The detailed information and display windows were shown in Supplementary Material S2.
Discussion
To our knowledge, this is the first study to evaluate whether the optimal dose of β-lactam antibiotics in neonatal sepsis can be predicted using an ML-based CDSS. In recent years, ML analytics have become increasingly popular for individualized treatment, with application flexibility and convenience being its main advantages.48,49 However, building an accurate and precise model that can handle the intricacy of the interactions between features and the output value calls for large datasets. This restricts their ability to advance in the neonatal field. The neonatal β-lactam antibiotics application field faces a double challenge, as the actual %fT > MIC cannot be measured directly. A theoretical approach is to simulate datasets, build ML models based on these simulations, and validate them using external data.50
According to the published ML recommendations, our study is of high quality, having completed validation based on real-world data.50 The PopPK model utilizes a compartment structure as its framework and incorporates fixed and random effects. ML is a data-driven approach that eliminates mechanistic assumptions. The objective of ML models is accuracy. Although the ML model in this study was constructed based on simulated data (generated by the PopPK model), because of different mechanisms, the ML model may perform differently when confronted with new data and situations compared with the PopPK model.50 As shown by the real-world data validation results, the overall prediction accuracy, precision, recall, and F1-Score of the CDSS improved by 10.7%, 22.1%, 64.2, and 43.1%, respectively.
The PD target of time-dependent β-lactam antibiotics is the %fT > MIC. A dosing interval of at least 40–50% fT > MIC was accepted for adults. To prevent the development of antibiotic resistance and ensure effectiveness, neonates should be considered in an immunocompromised state and require a higher %fT > MIC. However, the specific target PD value in neonates remains controversial. Various PD targets range from 40% fT > MIC up to 100% fT > 4–5 x MIC in most neonatal studies.51, 52, 53, 54, 55, 56 There is no generally accepted PD target because it is usually defined based on expert opinions rather than theoretical arguments or in vitro data.57,58 Thus, in this study, various PD targets, 50%, 70%, and 100%, which were the three most common PD targets, were selected for subsequent ML model training and prediction. This made the final CDSS more flexible and useful for clinical applications.
The identification of infectious pathogens and their corresponding MICs are crucial for selecting appropriate antibiotic therapies. However, because it is difficult to obtain positive culture and antimicrobial susceptibility results for bacteria in newborns,59,60 empiric therapy is widely used to treat neonatal infectious diseases without information about the pathogens. In this study, according to national and international references, we determined the MIC values of the five most common pathogens of neonatal sepsis and selected one of each to cover all pathogens.
Several studies have utilized ML to predict antibiotic dosages. For example, Codde C et al. successfully predicted the steady-state AUC of daptomycin by incorporating peak concentrations, trough concentrations (C0), and other covariates for dose adjustments.61 Similarly, Ma et al. predicted trough concentrations of teicoplanin to optimize dosing regimens.62 Various studies, such as those by Bououda M et al.63 and Huang et al.,64 have focused on predicting steady-state AUC or trough concentrations of vancomycin. However, these studies primarily concentrate on adult populations, with limited attention to neonatal groups. Clinically, these studies are mainly used for dose optimization based on existing drug concentrations, with minimal research on initial dosing or comprehensive dose optimization throughout treatment. Mechanistically, existing ML studies mainly focus on concentration-dependent antibiotics and recommend doses based on AUC/C0 predictions, neglecting research on time-dependent antibiotics that require assessing the duration free drug concentrations remain above the MIC.
The high scores on the test set are attributed to the fact that our predictive model was built based on simulated data, whereas the external validation data were derived from the real world, making its conditions more complex. We understand and acknowledge that the validation set results may show a decrease compared to the test set, which is expected because as the validation set better simulates the complexity and uncertainty of the real world. When the model is applied in clinical settings, continuous incorporation of real data will lead to ongoing improvements, allowing the model to better adhere to the principles of “learning and confirming”. Overall, although there are differences between the predictive results of the validation set and test sets, they remain satisfactory and outperform traditional or alternative predictive methods.
The CDSS showed acceptable prediction performance. Using real-world data, all five drugs had a prediction accuracy of 80% or more. Initially, the CDSS has the potential to treat new neonatal patients. The clinician entered the target drug, the characteristics of the neonate (e.g., age, weight, biochemical information), and the PD target (50%, 70%, or 100% fT > MIC) into each module of the system, and the system automatically determines whether the concentration at the planned dose regimen would reach the target (individual concentrations above the default MIC). The clinician selects an appropriate initial dosing regimen based on the predicted results. Based on the clinical outcomes of the current dosing regimen, physicians may choose to implement further dose adjustments based on the CDSS. If the treatment is ineffective (excluding cases of bacterial resistance, in which case it is recommended to change the drug), the clinician can determine a new dose by adjusting the three modules: patient, dosage, and PD target; if the treatment is effective, the current dose is maintained. The application scenario for the CDSS was shown in Fig. 4.
Fig. 4.
Schematic diagram to illustrate how the clinical decision support system would be applied in clinical practice. CDSS: Clinical Decision Support System; PD: pharmacodynamics; CW: current weight; BW: birth weight; GA: gestational age; PNA: postnatal age; PMA: postmenstrual age; CREA: serum creatinine.
Limitations of the Study. Covariates related to disease status and concomitant medications were not included in this study. The clinical value of this dose selection system needs to be validated in clinical practice. The clinical dose selection system requires continual refinement in clinical practice to adhere to the principles of “learning and confirming.”
Conclusions
An ML-based clinical decision support system was successfully constructed to assist clinicians in making optimal dose decisions for β-lactam antibiotics to treat neonatal sepsis.
Contributors
Bo-Hao Tang wrote the manuscript; Wei Zhao designed the research; Bu-Fan Yao, Yue-E Wu, Guo-Xiang Hao, Wei Zhang, Xin-Fang Zhang, Shu-Meng Fu, Yi Zheng, Yue Zhou and John van den Anker performed the research; Bo-Hao Tang, Yue-E Wu, De-Qing Sun, Gang Liu and Guo-Xiang Hao analyzed the data. Bo-Hao Tang and Bu-Fan Yao directly accessed and verified the underlying data reported in the manuscript. Bo-Hao Tang and Wei Zhao were responsible for the decision to submit the manuscript. All authors read and approved the final version of the manuscript, and ensure it is the case.
Data sharing statement
The dataset collected for this study will be made available to others upon reasonable request. These data will be available upon publication of this study. Data will be made available through the email address of the corresponding author. Requests for data access must be approved by the relevant ethics boards and data custodians. Access will be granted based on a research proposal.
Declaration of interests
All authors declare that they have no competing interests.
Acknowledgements
This work was supported by the National Key Research and Development Program of China (2023YFC2706100); National Natural Science Foundation of China (82173897); Distinguished Young and Middle-aged Scholar of Shandong University; Innovation and Development Joint Fund of Natural Science Foundation of Shandong Province (ZR2022LSW007) and Natural Science Foundation of Shandong Province (ZR2022QH004). Authors were not precluded from accessing data in the study, and they accept responsibility to submit for publication.
Footnotes
Supplementary data related to this article can be found at https://doi.org/10.1016/j.ebiom.2024.105221.
Appendix ASupplementary data
References
- 1.Odabasi I.O., Bulbul A. Neonatal sepsis. Sisli Etfal Hastan Tip Bul. 2020;54(2):142–158. doi: 10.14744/SEMB.2020.00236. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 2.Youngster I., Avorn J., Belleudi V., et al. Antibiotic use in children - a cross-national analysis of 6 countries. J Pediatr. 2017;182:239–244. doi: 10.1016/j.jpeds.2016.11.027. [DOI] [PubMed] [Google Scholar]
- 3.de Hoog M., Mouton J.W., van den Anker J.N. New dosing strategies for antibacterial agents in the neonate. Semin Fetal Neonatal Med. 2005;10(2):185–194. doi: 10.1016/j.siny.2004.10.004. [DOI] [PubMed] [Google Scholar]
- 4.Wang J., Kumar S.S., Sherwin C.M., et al. Renal clearance in newborns and infants: predictive performance of population-based modeling for drug development. Clin Pharmacol Ther. 2019;105(6):1462–1470. doi: 10.1002/cpt.1332. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 5.Kou C., Li D.F., Tang B.H., et al. Clinical utility of a model-based amoxicillin dosage regimen in neonates with early-onset sepsis. Br J Clin Pharmacol. 2022;88(11):4950–4955. doi: 10.1111/bcp.15521. [DOI] [PubMed] [Google Scholar]
- 6.Wu Y.E., Wang T., Yang H.L., et al. Population pharmacokinetics and dosing optimization of azlocillin in neonates with early-onset sepsis: a real-world study. J Antimicrob Chemother. 2020;76(3):699–709. doi: 10.1093/jac/dkaa468. [DOI] [PubMed] [Google Scholar]
- 7.Wu Y.E., Hou S.S., Fang Z.Y., et al. Clinical utiliy of a model-based piperacillin dose in neonates with early-onset sepsis. Br J Clin Pharmacol. 2022;88(3):1179–1188. doi: 10.1111/bcp.15058. [DOI] [PubMed] [Google Scholar]
- 8.Zhao Y., Yao B.F., Kou C., Xu H.Y., Zhao W. Developmental population pharmacokinetics and dosing optimization of cefepime in neonates and young infants. Front Pharmacol. 2020;11:14. doi: 10.3389/fphar.2020.00014. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 9.Wu Y.E., Kou C., Li X., et al. Developmental population pharmacokinetics-pharmacodynamics of meropenem in Chinese neonates and young infants: dosing recommendations for late-onset sepsis. Children. 2022;9(12):1998. doi: 10.3390/children9121998. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 10.Wang Y.K., Wu Y.E., Li X., et al. Optimal dosing of ceftriaxone in infants based on a developmental population pharmacokinetic-pharmacodynamic analysis. Antimicrob Agents Chemother. 2020;64(11) doi: 10.1128/AAC.01412-20. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 11.Jacqz-Aigrain E., Leroux S., Thomson A.H., et al. Population pharmacokinetic meta-analysis of individual data to design the first randomized efficacy trial of vancomycin in neonates and young infants. J Antimicrob Chemother. 2019;74(8):2128–2138. doi: 10.1093/jac/dkz158. [DOI] [PubMed] [Google Scholar]
- 12.Tang B.H., Zhang J.Y., Allegaert K., et al. Use of machine learning for dosage individualization of vancomycin in neonates. Clin Pharmacokinet. 2023;62(8):1105–1116. doi: 10.1007/s40262-023-01265-z. [DOI] [PubMed] [Google Scholar]
- 13.Zhou J., Jiang L., Zhang Z.L., et al. Population pharmacokinetics and dosing optimization of mezlocillin in neonates and young infants. J Antimicrob Chemother. 2022;77(8):2238–2244. doi: 10.1093/jac/dkac176. [DOI] [PubMed] [Google Scholar]
- 14.Shi Z.R., Chen X.K., Tian L.Y., et al. Population pharmacokinetics and dosing optimization of ceftazidime in infants. Antimicrob Agents Chemother. 2018;62(4) doi: 10.1128/AAC.02486-17. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 15.Carmine A.A., Brogden R.N., Heel R.C., Speight T.M., Avery G.S. Cefotaxime. A review of its antibacterial activity, pharmacological properties and therapeutic use. Drugs. 1983;25(3):223–289. doi: 10.2165/00003495-198325030-00001. [DOI] [PubMed] [Google Scholar]
- 16.Moon Y.S., Chung K.C., Gill M.A. Pharmacokinetics of meropenem in animals, healthy volunteers, and patients. Clin Infect Dis. 1997;24(Suppl 2):S249–S255. doi: 10.1093/clinids/24.supplement_2.s249. [DOI] [PubMed] [Google Scholar]
- 17.Carmine A.A., Brogden R.N., Heel R.C., Romankiewicz J.A., Speight T.M., Avery G.S. Moxalactam (latamoxef). A review of its antibacterial activity, pharmacokinetic properties and therapeutic use. Drugs. 1983;26(4):279–333. doi: 10.2165/00003495-198326040-00001. [DOI] [PubMed] [Google Scholar]
- 18.van den Anker J.N. Population pharmacokinetics and dosing of amoxicillin in (pre)term neonates. Ther Drug Monit. 2006;28(6):816. doi: 10.1097/01.ftd.0000249946.20294.59. author reply -7. [DOI] [PubMed] [Google Scholar]
- 19.Munshi R., Talele G., Shah R. In-vitro evaluation of antimicrobial activities of Escherichia coli, Klebsiella pneumoniae, Salmonella typhi, Neisseria gonorrhoeae, and Candida albicans nosodes. Homeopathy. 2022;111(1):42–48. doi: 10.1055/s-0041-1727149. [DOI] [PubMed] [Google Scholar]
- 20.Jones R.N., Barry A.L., Thornsberry C., et al. Ceftazidime, a pseudomonas-active cephalosporin: in-vitro antimicrobial activity evaluation including recommendations for disc diffusion susceptibility tests. J Antimicrob Chemother. 1981;8(Suppl B):187–211. doi: 10.1093/jac/8.suppl_b.187. [DOI] [PubMed] [Google Scholar]
- 21.Hartman S.J.F., Upadhyay P.J., Hagedoorn N.N., et al. Current ceftriaxone dose recommendations are adequate for most critically ill children: results of a population pharmacokinetic modeling and simulation study. Clin Pharmacokinet. 2021;60(10):1361–1372. doi: 10.1007/s40262-021-01035-9. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 22.Leroux S., Roue J.M., Gouyon J.B., et al. A population and developmental pharmacokinetic analysis to evaluate and optimize cefotaxime dosing regimen in neonates and young infants. Antimicrob Agents Chemother. 2016;60(11):6626–6634. doi: 10.1128/AAC.01045-16. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 23.Joly-Guillou M.L., Kempf M., Cavallo J.D., et al. Comparative in vitro activity of Meropenem, Imipenem and Piperacillin/tazobactam against 1071 clinical isolates using 2 different methods: a French multicentre study. BMC Infect Dis. 2010;10:72. doi: 10.1186/1471-2334-10-72. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 24.Quan J.J., Wang Y., Ji J.S., et al. [The activity of moxalactam against Enterobacteriaceae and anaerobia in vitro] Zhonghua Yixue Zazhi. 2016;96(18):1459–1464. doi: 10.3760/cma.j.issn.0376-2491.2016.18.015. [DOI] [PubMed] [Google Scholar]
- 25.Varma S., Simon R. Bias in error estimation when using cross-validation for model selection. BMC Bioinf. 2006;7:91. doi: 10.1186/1471-2105-7-91. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 26.Krstajic D., Buturovic L.J., Leahy D.E., Thomas S. Cross-validation pitfalls when selecting and assessing regression and classification models. J Cheminformatics. 2014;6:10. doi: 10.1186/1758-2946-6-10. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 27.Dorogush A.V., Ershov V., Gulin A. CatBoost: gradient boosting with categorical features support 2018. https://arxiv.org/pdf/1810.11363.pdf Available from:
- 28.Prokhorenkova L., Gusev G., Vorobev A., Dorogush A.V., Gulin A., editors. CatBoost: unbiased boosting with categorical features. 2017. [Google Scholar]
- 29.Ogami C., Tsuji Y., Seki H., et al. An artificial neural network-pharmacokinetic model and its interpretation using Shapley additive explanations. CPT Pharmacometr Syst Pharmacol. 2021;10(7):760–768. doi: 10.1002/psp4.12643. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 30.Sharland M. Manual of childhood Infections: the blue Book. 2016. [Google Scholar]
- 31.Committee JF. British national formulary for children 2014-2015. 2014. [Google Scholar]
- 32.Guilhaumou R., Benaboud S., Bennis Y., et al. Optimization of the treatment with beta-lactam antibiotics in critically ill patients-guidelines from the French Society of Pharmacology and Therapeutics (Societe Francaise de Pharmacologie et Therapeutique-SFPT) and the French Society of Anaesthesia and Intensive Care Medicine (Societe Francaise d'Anesthesie et Reanimation-SFAR) Crit Care. 2019;23(1):104. doi: 10.1186/s13054-019-2378-9. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 33.Li X., Qi H., Jin F., et al. Population pharmacokinetics-pharmacodynamics of ceftazidime in neonates and young infants: dosing optimization for neonatal sepsis. Eur J Pharm Sci. 2021;163 doi: 10.1016/j.ejps.2021.105868. [DOI] [PubMed] [Google Scholar]
- 34.Smith P.B., Cohen-Wolkowiez M., Castro L.M., et al. Population pharmacokinetics of meropenem in plasma and cerebrospinal fluid of infants with suspected or complicated intra-abdominal infections. Pediatr Infect Dis J. 2011;30(10):844–849. doi: 10.1097/INF.0b013e31822e8b0b. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 35.Qi H., Kou C., Qi Y.J., et al. Population pharmacokinetics and dosing optimization of latamoxef in neonates and young infants. Int J Antimicrob Agents. 2019;53(3):347–351. doi: 10.1016/j.ijantimicag.2018.11.017. [DOI] [PubMed] [Google Scholar]
- 36.Tang B.H., Wu Y.E., Kou C., et al. Population pharmacokinetics and dosing optimization of amoxicillin in neonates and young infants. Antimicrob Agents Chemother. 2019;63(2) doi: 10.1128/AAC.02336-18. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 37.Bizzarro M.J., Raskind C., Baltimore R.S., Gallagher P.G. Seventy-five years of neonatal sepsis at Yale: 1928-2003. Pediatrics. 2005;116(3):595–602. doi: 10.1542/peds.2005-0552. [DOI] [PubMed] [Google Scholar]
- 38.Vergnano S., Menson E., Kennea N., et al. Neonatal infections in England: the NeonIN surveillance network. Arch Dis Child Fetal Neonatal Ed. 2011;96(1):F9–F14. doi: 10.1136/adc.2009.178798. [DOI] [PubMed] [Google Scholar]
- 39.Stoll B.J., Hansen N.I., Sanchez P.J., et al. Early onset neonatal sepsis: the burden of group B Streptococcal and E. coli disease continues. Pediatrics. 2011;127(5):817–826. doi: 10.1542/peds.2010-2217. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 40.Muller-Pebody B., Johnson A.P., Heath P.T., et al. Empirical treatment of neonatal sepsis: are the current guidelines adequate? Arch Dis Child Fetal Neonatal Ed. 2011;96(1):F4–F8. doi: 10.1136/adc.2009.178483. [DOI] [PubMed] [Google Scholar]
- 41.Resch B., B R., N H. Comparison between pathogen associated laboratory and clinical parameters in early-onset sepsis of the newborn. Open Microbiol J. 2016;10:133–139. doi: 10.2174/1874285801610010133. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 42.Li-yuan G. CNKI: Chongqing Medical University; 2018. Literature research on the pathogen of early-onset sepsis in mainland China. [Google Scholar]
- 43.Li Y.P., Zhu Y.X., Wu T., et al. Analysis of the distribution of late-onset sepsis pathogens and the influence factors of late-onset sepsis on very low and extremely low birth weight infants. Chinese J Gen Pract. 2018;16(11):3. [Google Scholar]
- 44.Duan C., Hu Y.L., Qin X.L., et al. Distribution and drug sensitivity analysis of pathogens isolated from early － onset and late － onset neonatal sepsis. Chin J Health Lab Technol. 2019;29(7):3. [Google Scholar]
- 45.Wang G.Y., Chen Y.G., Zhou B., et al. Analysis on distribution and drug resistance of clinical pathogens of neonatal septicemia and its treatment strategies. Anti Infect Pharm. 2019;16(4) [Google Scholar]
- 46.Wei-xin Z. Clinical characteristics and pathogenic bacteria of early-onset and late-onset neonatal sepsis. J Guangdong Med Univ. 2015;35 [Google Scholar]
- 47.Liu C.X., Jing W.L., Chen X., et al. Clinical characteristics and pathogens of early and late neonatal sepsis analysis. Modern Instr Med Treat. 2016;22 [Google Scholar]
- 48.Woillard J.B., Labriffe M., Debord J., Marquet P. Tacrolimus exposure prediction using machine learning. Clin Pharmacol Ther. 2021;110(2):361–369. doi: 10.1002/cpt.2123. [DOI] [PubMed] [Google Scholar]
- 49.Tang B.H., Guan Z., Allegaert K., et al. Drug clearance in neonates: a combination of population pharmacokinetic modelling and machine learning approaches to improve individual prediction. Clin Pharmacokinet. 2021;60(11):1435–1448. doi: 10.1007/s40262-021-01033-x. [DOI] [PubMed] [Google Scholar]
- 50.Li Q.Y., Tang B.H., Wu Y.E., et al. Machine learning: a new approach for dose individualization. Clin Pharmacol Ther. 2023;115(4):727–744. doi: 10.1002/cpt.3049. [DOI] [PubMed] [Google Scholar]
- 51.Pullen J., Driessen M., Stolk L.M., et al. Amoxicillin pharmacokinetics in (preterm) infants aged 10 to 52 days: effect of postnatal age. Ther Drug Monit. 2007;29(3):376–380. doi: 10.1097/FTD.0b013e318067de5c. [DOI] [PubMed] [Google Scholar]
- 52.Pullen J., Stolk L.M., Nieman F.H., Degraeuwe P.L., van Tiel F.H., Zimmermann L.J. Population pharmacokinetics and dosing of amoxicillin in (pre)term neonates. Ther Drug Monit. 2006;28(2):226–231. doi: 10.1097/01.ftd.0000198648.39751.11. [DOI] [PubMed] [Google Scholar]
- 53.De Cock R.F., Smits A., Allegaert K., et al. Population pharmacokinetic modelling of total and unbound cefazolin plasma concentrations as a guide for dosing in preterm and term neonates. J Antimicrob Chemother. 2014;69(5):1330–1338. doi: 10.1093/jac/dkt527. [DOI] [PubMed] [Google Scholar]
- 54.Shang Z.H., Wu Y.E., Lv D.M., et al. Optimal dose of cefotaxime in neonates with early-onset sepsis: a developmental pharmacokinetic model-based evaluation. Front Pharmacol. 2022;13 doi: 10.3389/fphar.2022.916253. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 55.Burstein A.H., Wyble L.E., Gal P., et al. Ticarcillin-clavulanic acid pharmacokinetics in preterm neonates with presumed sepsis. Antimicrob Agents Chemother. 1994;38(9):2024–2028. doi: 10.1128/aac.38.9.2024. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 56.Adrianzen Vargas M.R., Danton M.H., Javaid S.M., et al. Pharmacokinetics of intravenous flucloxacillin and amoxicillin in neonatal and infant cardiopulmonary bypass surgery. Eur J Cardio Thorac Surg. 2004;25(2):256–260. doi: 10.1016/j.ejcts.2003.11.017. [DOI] [PubMed] [Google Scholar]
- 57.Roberts J.A., Paul S.K., Akova M., et al. DALI: defining antibiotic levels in intensive care unit patients: are current beta-lactam antibiotic doses sufficient for critically ill patients? Clin Infect Dis. 2014;58(8):1072–1083. doi: 10.1093/cid/ciu027. [DOI] [PubMed] [Google Scholar]
- 58.Scharf C., Liebchen U., Paal M., et al. The higher the better? Defining the optimal beta-lactam target for critically ill patients to reach infection resolution and improve outcome. J Intensive Care. 2020;8(1):86. doi: 10.1186/s40560-020-00504-w. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 59.Blackburn R.M., Muller-Pebody B., Planche T., et al. Neonatal sepsis - many blood samples, few positive cultures: implications for improving antibiotic prescribing. Arch Dis Child Fetal. 2012;97(6):F487–F488. doi: 10.1136/archdischild-2012-302261. [DOI] [PubMed] [Google Scholar]
- 60.Ng P.C., Ma T.P.Y., Lam H.S. The use of laboratory biomarkers for surveillance, diagnosis and prediction of clinical outcomes in neonatal sepsis and necrotising enterocolitis. Arch Dis Child Fetal. 2015;100(5):F448–F452. doi: 10.1136/archdischild-2014-307656. [DOI] [PubMed] [Google Scholar]
- 61.Codde C., Rivals F., Destere A., et al. A machine learning approach to predict daptomycin exposure from two concentrations based on Monte Carlo simulations. Antimicrob Agents Chemother. 2024;68(5) doi: 10.1128/aac.01415-23. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 62.Ma P., Liu R., Gu W., et al. Construction and interpretation of prediction model of teicoplanin trough concentration via machine learning. Front Med. 2022;9 doi: 10.3389/fmed.2022.808969. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 63.Bououda M., Uster D.W., Sidorov E., et al. A machine learning approach to predict interdose vancomycin exposure. Pharm Res. 2022;39(4):721–731. doi: 10.1007/s11095-022-03252-8. [DOI] [PubMed] [Google Scholar]
- 64.Huang X., Yu Z., Bu S., et al. An ensemble model for prediction of vancomycin trough concentrations in pediatric patients. Drug Des Dev Ther. 2021;15:1549–1559. doi: 10.2147/DDDT.S299037. [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.