Optimized explainable AI and digital twin for patient flow improvement in ICU during respiratory epidemics

Abstract

Background

Respiratory epidemics often place substantial pressure on intensive care units (ICU), which are continuously challenged to managing acute and life-threatening conditions under unpredictable workloads. During these periods, ICUs usually exhibit inefficient patient flows, treatment delays, and critical resource shortages. Proactive decision-making and precise interventions are therefore pivotal for patient survival and minimizing long-term sequelae.

Methods

This paper proposes a robust approach combining Artificial Intelligence (AI), Bayesian Optimization, and Digital Twin (DT) to support ICU patient flow management. An eXtreme Gradient Boosting (XGBoost) algorithm is used to predict the patient transfer probability from the emergency department (ED) to the ICU within the next 24 h. Bayesian optimization is employed for efficient hyperparameter tuning of the XGBoost model. Then, the transfer predictions are inserted into a DT to verify ICU capacity for timely care and design interventions for process mismatches.

Results

A case study from a European healthcare group validates the proposed approach. The specificity of the prediction XGBoost model was 94.90% (CI 95% 91.72% − 97.11%), whereas the sensitivity was 81.55% (CI 95% 72.70% − 88.51%). Finally, the median ICU bed waiting time decreased to between 66.74 and 69.38 h after implementing a patient transfer policy with a partner hospital having available ICU beds.

Conclusions

This study demonstrates the effectiveness of AI-DT in predicting the probability of ICU transfers, assessing the operational response of emergency wards and intensive care units, and crafting practical scenarios for enhancing patient flow management.

Background

Respiratory epidemics are characterized by abrupt increases in the incidence of severe disease, which impose disproportionate pressure on emergency departments (EDs) and intensive care units (ICUs). During these surges, critical care systems must concurrently ensure the timely escalation of care for high-acuity patients while functioning under conditions of extreme uncertainty, constrained capacity, and limited specialized resources. In practice, this tension often manifests as impaired patient flow between EDs and ICUs, prolonged waiting times, ED boarding, and shortages of beds, ventilators, and trained personnel. These operational disruptions have direct clinical consequences, as delayed in accessing intensive care are consistently associated with increased morbidity, mortality, and long-term functional impairments.

In response to these challenges, artificial intelligence (AI) has been increasingly investigated as a tool to support the early identification of patients who are likely to require ICU admission or transfer from the ED. Predictive models that draw routinely collected clinical data, including triage variables, physiological measurements, laboratory results, comorbidity profiles, and prior healthcare utilization, have demonstrated encouraging performance in retrospective and prospective evaluations. Recent evidence shows that AI-based models can reliably forecast the ICU admission risk and foresee near-term resource demands, underscoring their potential as real-time decision-support instruments in acute care environments [1].

The expansion of AI in healthcare extends well beyond critical care. In recent years, deep learning architectures have demonstrated remarkable diagnostic accuracy in medical imaging tasks from breast tumor segmentation in ultrasound [2] to the automated detection of tuberculosis in chest radiographs [3]. Complementary advances in transfer learning and large language models have further broadened the reach of AI, enabling effective clinical decision support even in settings where labeled data is scarce [4]. Alongside these developments, the integration of AI with emerging technologies such as blockchain and the Internet of Medical Things has opened new possibilities for personalized monitoring and resource optimization across high complexity care settings [5]. Taken in context, these advances reflect a broader shift in which AI is no longer confined to isolated diagnostic tasks but is increasingly positioned as a foundation for operational decision-making across the healthcare continuum a trajectory that directly motivates the present study.

Despite these advances, the translation of predictive accuracy into operational benefit remains limited. There are two recurrent barriers hindering current implementations. First, prediction outputs are often treated as independent risk estimates, without being embedded into dynamic operational workflows or decision-making processes. Second, the performance and reliability of these models can be undermined by institutional heterogeneity and temporal shifts in patient populations, particularly during epidemic phases. Consequently, a substantial gap persists between accurate individual level predictions and actionable interventions that can improve patient flow during times of severe capacity constraints.

Beyond individual risk stratification, health systems require anticipatory insights at the system level, including short- and medium-term forecasts of ICU and hospital occupancy. In the wake of the COVID-19 pandemic, several scalable and modular forecasting frameworks were developed to project bed demand under evolving epidemic conditions [6, 7]. Complementary optimization-based approaches have also been proposed to support ICU bed allocation during large-scale outbreaks, underscoring the importance of aligning epidemiological trends with capacity planning decisions [8]. Although these methods contribute to strategic preparedness, they typically operate at an aggregate level and remain insufficiently integrated with operational micro processes transfers from the ED to the ICU, managing queue formation, and addressing bottlenecks.

Digital twins (DT) have emerged as a promising paradigm to bridge this gap to better represent hospital systems by providing continuously updated virtual counterparts that are synchronized with real-world data. Recent meta-analyses and scoping reviews document the rapid expansion of DT applications across various healthcare domains, while also identifying persistent challenges related to interoperability, data governance, validation, and clinical integration [9, 10]. A consistent conclusion across these reviews is that DTs achieve their greatest value when combined with predictive models, as well as simulation or optimization techniques. This combination enables healthcare organizations to evaluate alternative interventions and anticipate system-level failures before they occur.

In acute and critical care settings, DT approaches are particularly compelling because of strong coupling between patient flow, resource availability, and stochastic variability. Recent methodological advancements have proposed DT architectures that integrate discrete event simulation (DES) to optimize critical care workflows based on real-time system states. This provides a concrete foundation for DT-enabled operations in the ICU [11]. Additionally, parallel efforts in emergency care further illustrate the potential of DT-driven decision support to reduce waiting times and enhance capacity-aware management strategies [12].

A notable trend in the literature is the convergence of AI based prediction with simulation methodologies, particularly DES. This combination is used to identify flow bottlenecks, quantify queue dynamics, and evaluate alternative operational policies, such as prioritization rules, staffing adjustments, and discharge strategies. Evidence from ICU capacity management during the COVID 19 crisis indicates that integrated AI and simulation approaches can help mitigate mismatches between patient demand and available resources, supporting data-driven operational redesign [13]. More broadly, simulation-based bed capacity management frameworks that incorporate predictive analytics have been proposed to strengthen system resilience during large-scale health emergencies [14].

Among the various predictive modeling techniques, eXtreme Gradient Boosting (XGBoost) has gained widespread adoption in clinical prediction tasks because its strong performance with structured healthcare data. However, the effectiveness of XGBoost models is highly sensitive to hyperparameter selection. Conventional tuning strategies, such as manual adjustments or exhaustive grid searches, can often be inefficient and difficult to reproduce within tight time constraints. Bayesian Optimization has therefore emerged as an attractive alternative, offering a principled and computationally efficient approach to hyperparameter tuning. Recent healthcare applications show that Bayesian optimized XGBoost pipelines improve both predictive performance and model stability, reinforcing their suitability for time-sensitive clinical decision support systems [15].

The body of evidence reviewed points in a consistent direction: short horizon AI predictions, computationally efficient hyperparameter tuning, and simulation based digital twins are mature enough individually to support ICU management, yet they have rarely been deployed together. Most studies stop at prediction or at simulation, and fewer still make their outputs interpretable enough for clinicians to act on them with confidence. The gap, in practical terms, is the distance between knowing which patients are likely to need intensive care and having a tested, operational plan ready when they arrive. This study was built around that gap. Bringing together a Bayesian optimized XGBoost model, SHAP based clinical explanations, and a Digital Twin, we propose a framework that travels the full distance from individual patient risk to pretested interventions for managing ED to ICU flow during respiratory epidemics [5, 6].

The main contribution of this article is three-fold:

• i)Optimally predict the patient transfer probability to the ICU within the next 24 h after their entry into the EW. Bayesian optimization is used to fine-tune the hyperparameters of the XGBoost algorithm, ensuring the best hyperparameter mix for optimal prediction accuracy. The model is explainable via Shapley Additive Explanations (SHAP), which allows medical staff and hospital administrators to understand how individual features contribute to ICU transfer probability predictions.
• ii)Insert the XGBoost ICU transfer predictions into a DT emulating the EW-ICU pathway. This will allow for real-time evaluation of the balance between the anticipated volume of ICU admissions and the available capacity, within the next 24 h. With this information, healthcare administrators can determine if incoming patients will receive timely care in the ICU. If an imbalance is identified, they can make informed medical and operational decisions to address this situation.
• iii)Craft new operational scenarios uplifting patient flow management throughout the EW-ICU patient journey. The DT will be employed to pretest these scenarios before implementation, so that decision-makers to anticipatedly their effectiveness during a respiratory-disease outbreak.

This paper is organized as follows: The Methods section (Section “Methods”) outlines the study’s design, setting, and population, together with the strategies employed for feature selection, management of missing data, and defining outcomes. It also describes the data processing pipeline, the formulation of the predictive model, and the integration of the eXtreme Gradient Boosting (XGBoost) algorithm with Bayesian Optimization for hyperparameter tuning, in addition to the simulation environment used to construct the digital twin (DT). The Results section (Section “Results”) presents empirical findings from the case study, emphasizing the predictive performance of the Bayesian-optimized XGBoost model and the operational insights gained from DT simulations. The Discussion section (Section “Discussion”) interprets these findings in relation to existing literature, highlighting the methodological contributions, practical implications for managing ICU patient flow, and limitations that warrant further investigation. Finally, Section “Conclusions” concludes the paper and announces future works.

Methods

This research entails the integration of AI, Bayesian optimization, and digital twins for underpinning ICU patient flow management, especially during outbreaks of respiratory viral disease. The main objective is three-fold: (i) to optimally predict the patient transfer probability to the ICU within the next 24 h after their entry into the EW, (ii) to appraise the balance between the expected volume of ICU admissions within the same timeframe and the available capacity, and (iii) to craft operational scenarios uplifting patient flow management. Consequently, medical staff and hospital administrators will be able to transition from reactive to predictive foresight, underpinned by a user-friendly and effective decision-making platform useful for facing this challenge.

This project was undertaken with adherence to the principles outlined in the Declaration of Helsinki. Formal approval and informed consent were received from the Ethics Committee of the European healthcare group (Agreement number: 14-12-2021-004; Access request ID:39) given the project’s associated risks and context. Additionally, a confidentiality agreement was established to safeguard data privacy and confirm that the information will be employed exclusively for this research.

We retrospectively gathered sociodemographic and clinical data from respiratory-infected patients.

The inclusion criteria were the following:

• (i)Age between 21 and 100 years.
• (ii)At least two measurements of clinical vital signs: The first taken at the triage station and the second during the patient’s stay in the emergency ward (EW).
• (iii)Availability of D-dimmer concentration and C Reactive Protein (CRP) results.

The dataset, derived from the Hospital’s Patient Management software, comprised anonymized records from 4,479 patients who visited the showcased EW between February (2020) and February (2021). This season saw the presence of multiple respiratory viruses with high viral loads, including influenza, COVID-19, and Respiratory Syncitial Virus (RSV). Each patient record comprised the following information: a unique patient code, D-dimmer concentration, Lactate Dehydrogenase (LDH), CRP, Systolic arterial pressure (two readings), Diastolic arterial pressure (two readings), Oxygen saturation (two readings), Heart rate (two readings), Body temperature (two readings), Age, and Sex. Some features were measured twice as healthcare decisions are often reliant on the patient’s evolution over time [16–18]. More specifically, patients may go through different process stages (transitory status) hinging upon their response to medical treatment, disease progress, and context. Therefore, the use of time-dependent characteristics can guide doctors on deciding more accurately, which patient needs to be transferred to critical care. On the other hand, thoracic images were not taken into account in the dataset because of limited availability and varying analysis methods among physicians.

Missing data for each potentially predictive feature was imputed by employing the median value from the outcome variable layers within the cohort [13, 19]. This stratified imputation approach was employed to preserve the distributional differences between outcome groups and to avoid distortion of predictor–outcome relationships that may arise from global imputation strategies. Although this approach may introduce an optimistic bias, its impact was mitigated through a cross-validation strategy to ensure a reliable estimate of performance [20–22]. Additionally, we removed 1,717 patients due to inconsistent registrations of clinical lab features (Fig. 1). Extant records remain unchanged, and the derived dataset was utilized to train the patient transfer probability prediction model. It is also necessary to underscore the use of random undersampling to tackle the imbalance between the number of patients who were moved to ICU care and those who were not.

Fig. 1.

Data management procedure

Feature selection is conducted through an Analysis of Variance (ANOVA) (α = 0.05). Furthermore, feature importance rankings derived from the Gini coefficient highlight which variables contribute most to model accuracy.

We also defined a dicotomic response variable to denote the transfer need to intensive care for each patient admissioned in the EW. Specifically, a value of “1” was assigned to represent patients moved to ICU; whereas, “0” was used for those who did not require this type of assistance. The distance between these two numbers designates the transfer probability space where different cut-offs can be established for particular intervention levels. Hospital administrators can us this information to develop operational and medical plans to diminish this probability and uplift the preparedness of these units in anticipation of expected admissions.

Not less important is the partitioning scheme (70/30) employed to generate the training and test cohorts needed for adjusting and refining the XGBoost model. Each cohort contains patients with/without transfer requests. The training set was utilized to teach the XGBoost model to detect context-dependent patterns and learn the parameters observed in the two patient types. Conversely, the test set was employed to evaluate the model performance, which helps assess its generalizability and detect any overfitting. In other words, this set exhibits how well the XGBoost algorithm predict the ICU transfer probability.

EXtreme gradient boosting (XGBoost)

XGBoost is often used to build a fast, scalable, accurate, and end-to-end tree boosting model, effectively illustrating the relationship between predictors and a probabilistic response variable [23–26]. Building on this premise, the XGBoost is now used for forecasting the ICU transfer probability within the next 24 h after EW admission, based on significant predictors selected from the aforementioned dataset.

XGBoost has proven to address overfitting problems due to a regularized procedure that simplifies the parallel processing algorithm [27]. The far-reaching impact of XGBoost is also supported by its ability to face real-world high-scale problems utilizing a minimal quantity of resources, which is greatly relevant for the sustainability of hospitals [28, 29]. This algorithm has also demonstrated superior performance over other machine learning methods when processing medical data [29, 30]. Its advantages extend to the ability to handle datasets where the prevalence of certain healthcare events is small, thus preventing uniform predictions for all patients [31, 32]. This is critical in this context, considering the small proportion of patients moved to the ICU as indicated by the dataset’s descriptive statistics.

The mathematical formulation related to the XGBoost procedure is presented in Eq. (1) and Eq. (2). Here, denotes the ICU transfer probability for the patient i within the next 24 h after EW admission, while symbolizes the predictive value of the k-th tree for the patient i based on significant predictors . In Eq. (2) represents the prediction derived from the preceding experiment at time t.

1

2

The observation may be right censored as the patient could be discharged early or die within the next 24 h after EW admission. is the outcome variable and is defined by Eq. 3. The outcome timing is represented by Eq. 4.

3

4

The software R (v. 4.3.2) and the Integrated Development Environment Rstudio (v. 2023.09.1) were employed to develop the XGBoost model supported by a set of libraries, including: data.table (v. 1.17.8), xgboost (v. 1.7.8.1), caret (v. 6.0–94), dplyr (v. 1.1.4), ggplot2 (v. 4.0.1), MLmetrics (v. 1.1.3), pROC (v. 1.18.5), epiR (v. 2.0.66), SHAPforxgboost (v. 0.2.0), rBayesianOptimization (v. 1.2.2), and shapviz (v. 0.10.3).

Bayesian optimization uplifts this modelling procedure by fine-tuning the model’s hyperparameters [33–35]. This method is preferred over other Grid Search and Random Search when the function is very costly and time-consuming during evaluation. Its adaptative nature allows to minimize the number of iterations used to find the best set of hyperparameters [36, 37]. Furthermore, it uses a Gaussian process that not only predicts outcomes but also determines how accurate these predictions are within each search space. Likewise, Bayesian optimization strikes a balance between exploration and exploitation by employing an acquisition function, which prevents search from becoming focused solely on local optima.

To make the model more actionable for clinicians, our approach has integrated Shapley Additive Explanations (SHAP), an explainability technique that can provide a patient-specific breakdown of risk factors [38–40]. This integration enables healthcare providers to better understand and trust the model’s outputs, ultimately facilitating its adoption in clinical settings. Nonetheless, these contributions reflect the XGBoost model’s internal logic rather than causal relationships, and hence should not be interpreted as direct targets of intervention.

Digital twin (DT)

The patient path from the EW to ICU is mapped through a flow diagram. This illustration is based on direct observation and established medical protocols. Thereby, it is possible to identify the multiplicity of treatment options, process stages, transitory/absorbing states, as well as important variables and parameters involved.

Data associated with the variables are gathered to build the dataset underpinning the DT. The quality of data is crucial for ensuring the reliability of DT in assessing patient flow management, improving demand-capacity balance, and designing improvement scenarios [41, 42]. This data undergoes preprocessing using an input data analysis comprising three major steps: intra-variable independence analysis, homogeneity analysis, and goodness-of-fit testing. The intra-variable independence analysis is operationalized through a runs test (α = 0.01) that determines if a certain process variable is of random nature. The homogeneity analysis is then conducted via a Kruskal-Wallis test (α = 0.01) to identify any subgroups within the same dataset. This analysis helps assess whether the data should be separated based on specific drivers. The goodness-of-fit test will then define the probability distribution representing the behaviour of the process variables, either as single dataset or as multiple pipelines. For the latter, a stochastic distribution must be derived for each subgroup. The Kolmogorov-Smirnov test (α = 0.01) will assist in this task by identifying the distribution parameters to be inserted into the DT.

Afterwards, the modelers proceed with inserting all the information derived from the input data analysis and process mapping into a virtual real-time replica. The transfer predictions arising from the XGBoost algorithm are also inputted into the DT as a vector array. Digital Twins can provide a real-time representation of patient flows from the EW to ICU, underpinning fast and effective decision-making in complex scenarios, such as those posed by the respiratory-disease outbreaks. Aside from this pivotal aspect, several reasons bolster the DT usage in this context:

• i)It facilitates engagement with healthcare decision-makers by animating patient flows, resources, failures, and key performance indicators.
• ii)It allows decision-makers to evaluate how critical resources are being used.
• iii)It is versatile in creating and simulating different operational scenarios that may be pretested before implementation without incurring in the cost and timing waste often associated with trial-to-error approaches [43, 44]. This is highly useful in the intensive care context, where implementing non-effective strategies may jeopardize patients’ lives.

Despite the DT strengths, its use in patient flow management is very limited [45, 46]. The simulation software selected for creating the DT is ARENA ^® v. 16.2. The DT is statistically validated through a one-sample sign test (α = 0.01) to verify its comparability to the real-world healthcare system. If equivalent, a demand-capacity balance analysis is performed using the DT to discriminate whether the ICU is equipped to provide timely care for prioritized respiratory-affected patients. Otherwise, the virtual replica must be revised and adjusted until achieving similarity with the real-world system. In case of imbalance, new operational scenarios are modelled and simulated in the DT to determine if they will be effective for facilitating patient flow in the wild. The effectiveness of these scenarios is further tested by a Mann-Whitney test (α = 0.01). The key indicator for measuring patient flow management will be the median waiting time for an ICU bed. If the p-value is minor that the significance level, then the proposed scenario will be categorized as “effective” (); on the contrary, it will be rejected for implementation (). The pretesting procedure will verify the hypothesis shown in Eq. (5) where represents the current median waiting time for an ICU bed, whereas is the median ICU bed waiting time that the hospital would achieve if the proposed plan were put into action.

5

Results

Table 1 portrays the characteristics of 1,326 patients who arrived at the EW during a respiratory-disease season marked that featured multiple viruses, including influenza A (Notably H3N2), COVID-19 (SARS-CoV-2), and RSV. Patients classified as Group T were transported to the ICU, whereas those in Group NT did not need intensive care.

Table 1.

Descriptive statistics of quantitative candidate predictive features

Quantitative features	NT* (Mean ± SD) n6 = 928		T* (Mean ± SD) n5 = 398		p-value	Cohen’s d
SYSTOLIC_PRESSURE_1	130.95 ± 22.11		130.59 ± 15.32		0.764	-0.02
DIASTOLIC_PRESSURE_1	76.78 ± 28.25		76.19 ± 9.5		0.688	-0.03
TEMPERATURE_1	36.69 ± 0.8		36.97 ± 0.77		< 0.001	0.36
HEART_RATE_1	89.82 ± 17.45		91.8 ± 14.32		0.046	0.12
OXYGEN_SATURATION_1	93.1 ± 6.55		90.02 ± 8.58		< 0.001	-0.4
SYSTOLIC_PRESSURE_2	130.98 ± 22.42		131.26 ± 16.45		0.82	0.01
DIASTOLIC_PRESSURE_2	77.39 ± 34.76		76.45 ± 10.77		0.596	-0.04
TEMPERATURE_2	36.69 ± 0.81		36.98 ± 0.8		< 0.0001	0.36
HEART_RATE_2	89.75 ± 17.62		92.04 ± 15.34		0.025	0.14
OXYGEN_SATURATION_2	93.17 ± 5.93		89.46 ± 9.6		< 0.0001	-0.47
D.DIMER	1352.16 ± 3980.08		3074.77 ± 9811.89		< 0.001	0.23
LACTATE_DEHYDROGENASE	558.86 ± 219.68		750.35 ± 402.46		< 0.0001	0.59
C-REACTIVE-PROTEIN (CRP)	85.68 ± 73.47		130.08 ± 93.58		< 0.0001	0.53
AGE	66.65 ± 17.02		64.95 ± 15.25		0.086	-0.1
Qualitative features	M	F	M	F	p -value	OR
SEX	56.5%	43.5%	73.4%	26.6%	< 0.001	2.12

(*Group T were transported to the ICU, whereas those in Group NT did not need intensive care)

Values are presented as mean ± standard deviation (SD). P-values were obtained using ANOVA for quantitative variables and appropriate statistical tests for categorical variables. Effect size of numerical features was calculated using Cohen’s d. Interpretation of Cohen’s d: < 0.2 (trivial - no clinical relevance), 0.21–0.49 (small - possibly clinically relevant), 0.5–0.79 (medium – clinically meaningful) and ≥ 0.8 (large – high potential clinical relevance). The Odds Ratio (OR) was performed for measuring the effect size of Sex.

ANOVA was undertaken for supporting feature selection. The factors with significant influence on the ICU transfer probability were: Temperature_1 (p-value = < 0.0001; d = 0.36), Heart rate_1 (p-value = 0.046; d = 0.12), Oxygen saturation_1 (p-value = < 0.001; d = -0.4), Temperature_2 (p-value = < 0.0001; d = 0.36), Heart rate_2 (p-value = 0.025; d = 0.14), Oxygen saturation_2 (p-value = < 0.0001; d = − 0.47), D-dimmer (p-value = < 0.001; d = 0.23), LDH (p-value = < 0.0001; d = 0.59), CRP (p-value = < 0.0001; d = 0.53), and Sex (p-value = < 0.001; OR = 2.12).

It is good to emphasize that Oxygen saturation, body temperature, and heart rate are significant in both measurements. These time-dependent indicators are essential for monitoring patient health progress. Notably, the T-value decreases in the second Oxygen saturation report, suggesting a decline in patient’s respiratory function. The negative direction in T-values signifies that a reduction in oxygen saturation level is statistically related to an augmented ICU transfer probability. In fact, the second reading of this indicator is lower for ICU-transferred patients (µ_T2 = 89.46%; σ_T2 = 9.6%) when compared to non-ICU-transferred people (µ_NT2 = 93.17%; σ_T2 = 5.93%). There is also a slight mean decrease and more volatility in this metric when contrasting the first (µ_T1 = 90.02%; σ_T1 = 8.58%) and second (µ_T2 = 89.46%; σ_T2 = 9.6%) reports associated with ICU patients. A similar pattern was detected when comparing the first (µ_T1 = 91.8 beats; σ_T1 = 14.32 beats) and second (µ_T2 = 92.04 beats; σ_T2 = 15.34 beats) heart rate registrations in the ICU-transferred population. In this case, this variable increased and exhibited more instability, indicating a deterioration in health status. Likewise, the heart rate was concluded to be higher among the ICU cohort (µ_NT2 = 89.75 beats Vs. µ_T2 = 92.04 beats). The less significant differences between the two readings were pinpointed in body temperature, a finding supported by the almost null change of T-values. Nonetheless, this variable tended to be greater in the ICU-transferred group with an average difference of 0.29 °C compared to the group NT.

On the other hand, three blood indicators (CRP, D-dimmer, and LDH) were found to be significant on the ICU transfer probability. All the T-values for these features were positive, indicating that an increase in these indicators reflects poorer health conditions in patients and an augmented ICU likelihood admission. In fact, the gaps between groups T and NT are very high in terms of absolute average numbers: CRP (µ_NT = 85.68 mg/L Vs. µ_T = 130.08 mg/L), D-dimmer (µ_NT = 1352.16 ng/mL Vs. µ_T = 3074.77 ng/mL), and LDH (µ_NT = 558.86 U/L Vs. µ_T = 750.35 U/L). Albeit all these indicators are outside the normal ranges for adults in both groups, they demonstrate different stages of respiratory disease evolution. Of course, the most concerning context is the one presented in patients who were ultimately transferred to intensive care. Not less relevant is the variability of these measures. For LDH, the standard deviation in NT group approximately duplicates the one derived from group T (σ_NT = 219.68 U/L Vs. σ_T = 402.46 U/L). The highest variation change is observed in D-Dimmer where SD in the transferred cohort is more than twice that of the non-transferred patients (σ_NT = 3980.08 U/L Vs. σ_T = 9811.89 U/L). Moreover, the SD triplicates the mean D-dimmer concentration, which denotes a strong relationship with the respiratory disease evolution. The same behaviour was also noted in CRP, albeit to a lesser extent.

Finally, the T-value for Sex was negative signaling that men, coded as “1” in the dataset, are more likely to experience more severe complications than women (coded as “2”) when facing a respiratory disease. This is also glaring given that the proportion of men (73.4%) constitute a larger proportion of group T, compared to only 26.6% for women.

On a different tack, Fig. 2 displays a ranking of the features based on the Gini coefficient. This index highlights which variables most significantly contribute to the model’s accuracy. In this instance, the top five predictors are: D-dimmer (w = 1), CRP (w = 0.595), Systolic pressure_1 (w = 0.197), Diastolic pressure_1 (w = 0.177), and Age (w = 0.162). It is noteworthy that both D-dimmer and CRP are strong ICU transfer probability predictors.

Fig. 2.

Gini coefficients of candidate ICU transfer probability features

The optimal XGBoost hyperparameters emanating from Bayesian optimization are as follows: eta (0.09476265), max-depth (3), min_child_weight (1.49602911), subsample (0.76311331), and colsample_bytree (0.50610608). The model’s performance metrics denote a very strong discriminatory power, with all metric exceeding 81% (Table 2). Particularly, the accuracy (91.44% − 95% CI [88.24% − 94.00%] suggests that the XGBoost algorithm can correctly forecast between 88.24% and 94.00% the ICU transfer requirement in respiratory-affected patients. In contrast, the sensitivity (81.55% − 95% CI [72.70% − 88.51%] indicates that the XGBoost model will effectively identify between 72.70% and 88.51% of the patients needing intensive care. Simultaneously, the specificity (94.90% − 95% CI [91.72% − 97.11%] declares that the algorithm will accurately distinguish between 91.72% and 97.11% of cases where intensive care is not crucial for enhancing the patient’s health status. Similarly, the Area Under Receiver Operator Characteristic Curve (AUC-ROC) (Fig. 3) (95.92% − 95% CI [93.90% − 97.90%] evidences the excellent ability to discriminate between patients who need ICU transportation and those who do not. Additionally, McNemar’s test (χ² = 0.265; p-value = 0.6069) suggests a balanced distribution between false positives and false negatives. Therefore, there is evidence to claim that the model’s misclassification errors are symmetric (no systematic bias). In other words, the model is not skewed toward over-admitting patients to ICU or missing ICU cases.

Table 2.

Performance indicators of the XGBoost model (95% confidence level)

Accuracy (%)	Sensitivity (%)	Specificity (%)	Positive Predictive Value (%)	Negative Predictive Value (%)	AUC - ROC (%) (DeLong)	McNemar’s Test P-Value
91.44 (88.24-94.00)	81.55 (72.70-88.51)	94.90 (91.72–97.11)	84.85 (76.24–91.26)	93.62 (90.22–96.11)	95.92 (93.90–97.90)	0.6069

Fig. 3.

Area under receiver operator characteristic curve for the ICU transfer probability prediction model

Additionally, repeated 5-fold cross-validation was conducted yielding a mean AUC of 0.9566 [95% CI(0.9417–0.9691)], indicating excellent discriminative performance and low variability across resampling iterations. These results suggest that the model is both accurate and stable. Furthermore, the probability threshold for classification was determined using the Youden Index derived from the ROC curve. The optimal probability threshold determined using the Youden Index was 0.43, yielding a sensitivity of 0.864 and a specificity of 0.935, indicating a well-balanced classification performance. A threshold of 0.43 indicates that patients with a predicted probability equal to or greater than 43% are classified as requiring ICU admission.

The calibration plot (Fig. 4) shows the relationship between predicted probabilities and observed outcomes across deciles of risk. The calibration slope and intercept were 1.075 and 0.131, respectively.

Fig. 4.

Calibration plot of the XGBoost model

Shapley Additive Explanations (SHAP) were calculated for all the patients considered in this study. Thereby, healthcare administrators, ICU supervisors, and medical staff can count on a breakdown of risk factors describing how they contribute to the ICU transfer prediction. Figure 5 provides a waterfall plot illustrating a patient with a high ICU transfer probability (0.927). The yellow bars represent the factors increasing the model prediction values, while the red bars denote features that push the final prediction to a lower ICU requirement probability. In this particular case, the patient’s CRP (174 mg/L) is markly elevated with reduced oxygen saturation level (89%), and older age (68 years old) are associated with higher predicted risk according to the model. Conversely, patient’s second systolic blood pressure reading is normal (100 mmHg), contributing to lowering the ICU transfer probability. These contributions should be interpreted as reflecting how the XGBoost model integrates input features to generate the prediction, rather than as proof of direct causal effects on ICU admission.

Fig. 5.

Breakdown of risk factors for a patient with an ICU transfer probability of 0.927

Sensitivity analysis

Undersampling vs. Oversampling

To address the class imbalance inherent in the dataset, particularly in the identification of patients requiring ICU admission, different sampling strategies were implemented and comparatively evaluated. We specifically compared the selected undersampling technique with Synthetic Minority Over-sampling Technique (SMOTE). SMOTE generates synthetic instances of the minority class based on feature space similarities. This method aims to enhance the representation of clinically relevant minority cases without discarding existing data (n = 2,762), thereby preserving the underlying data distribution.

The comparison is particularly focused on capturing minority class patterns associated with ICU admission. Both sampling strategies were integrated within the same modeling framework to ensure consistency in training procedures and prevent evaluation bias. Figure 6 presents the comparison results in terms of the AUC. The DeLong test concludes that the balancing methods offer similar AUC (D = 0.0108; df = 982.89; p-value = 0.9914).

Fig. 6.

Area under receiver operator characteristic curve for the ICU transfer: SMOTE Vs Undersampling

Hyperparameters derived from Bayesian optimization

A sensitivity analysis of the hyperparameter search process was conducted based on the optimization history obtained from Bayesian Optimization, and following a Bayesian Optimization Weighted Sensitivity Analysis (BOWSA) inspired approach. Figure 7 presents the relative importance of the evaluated XGBoost hyperparameters, measured as the absolute correlation between each parameter and the model performance in terms of AUC.

Fig. 7.

Sensitivity analysis of XGBoost hyperparameters using Bayesian optimization

Among the evaluated hyperparameters, max_depth (maximum tree depth) shows the highest importance value (0.371), followed by eta (learning rate) (0.218). The parameters colsample_bytree (feature subsampling ratio per tree) (0.113) and subsample (instance subsampling ratio) (0.103) exhibit intermediate importance values, while min_child_weight (minimum sum of instance weight needed in a child) is the last ranked (0.031).

In addition, to analyze the behavior of the hyperparameter optimization process in greater depth, sensitivity plots were generated based on the optimization history obtained through Bayesian optimization. Figure 8 (a-e) shows the relationship between the selected hyperparameters and model performance (AUC). Figure 9 shows a surface plot combining the two most dominant drivers (max_depth and eta) considering optimized min_child_weight, subsample, and colsample_bytree. The optimal values of these parameters: max_depth (3) and eta (0.094763) fall into the global optimal region (0.75 ≤ AUC ≤ 1.0) highlighted in yellow. The diagonal contours indicate a significant interaction between these two hyperparameters. In summary, the model is conservative and regularized with controlled learning dynamics. Specifically, there is a compensated trade-off between learning rate and model complexity. Moreover, it exhibits small trees with meaningful selective splits. Also, stochasticity (subsample < 1) is properly dampened by the learning rate, while offering strong generalization capability. The curves demonstrate non-convex landscape, stable high-performance region, and nonlinear interactions. It is also evident that the model learns complex patterns progressively using simple diverse trees.

Fig. 8.

Sensitivity plots showing the relationship between hyperparameters and model performance (AUC): (a) eta (learning rate), (b) max_depth (maximum tree depth), (c) min_child_weight (minimum child weight), (d) subsample (instance subsampling ratio), (e) colsample_bytree (feature subsampling ratio)

Fig. 9.

Sensitivity surface plot combining max-depth and eta

Scenario-dependent threshold trade-offs

There are three clinical scenarios considered in this analysis (Fig. 10): Early warning (0.20 ≤ p_t ≤ 0.3), Balanced threshold (p_t = 0.43), and ICU scarcity (0.55 ≤ p_t ≤ 0.70). The early warning scenario avoids missing ICU cases and accepts 0.43 unnecessary admissions to avoid 1 missed case. On the other hand, the ICU scarcity context accepts 1 missed ICU case to evade 1.5 redundant visits.

Fig. 10.

Scenario analysis for ICU need

The ICU transfer probabilities obtained from the XGBoost prediction model were later imported as a vector array in a digital twin mimicking the patient journey from the Emergency Ward (EW) to the ICU. The vector array was attached to the model under the name “ICU transfer” which was allocated to the model entity (patient) through an Assign module. Then, the entity passed by a filter modelled via a Decision module (Type: 2-way by condition) where the transfer probability was compared to the optimal threshold (0.43). Patients with a predicted probability equal to or greater than this threshold were classified as requiring ICU admission. This denotes the exact clinical decision point at which prediction is generated which corresponds to one hour after EW admission. The rationale behind this period is that the lab turnaround time for some blood indicators (i.e. LDH, CRP) used by the XGBoost model. A map including the different stages of the healthcare provided to respiratory-affected patients is outlined in Fig. 11. This graph also indicates where the predictor’s data is collated and when the XGBoost model produces the forecast.

Fig. 11.

Journey of respiratory-affected patient within the hospital

Direct observation and medical protocols enabled the modelers to identify four key process variables to include in the DT: Time Between Admissions, Triage Classification Time, Length of Stay in EW, and Length of Stay in ICU. Data for each of these variables were gathered from the Hospital’s information sources and analyzed regarding randomness, homogeneity, and goodness-of-fit. The runs test (α = 0.01) was applied to appraise the intra-variable independence of each variable. The results, in terms of p-value and differences between observed (OR) and expected runs (ER) are summarized in Table 3. All variables were concluded to be random and can therefore be represented by a probability distribution (p-value > 0.048; OR – ER ≈ 0). We also reported the homogeneity analysis outcomes performed through Kruskal-Wallis tests (α = 0.01) (Table 3). The H-value, adjusted for ties, and p-values were computed to support the examination of potential subgroups within the data. Different layers were found in the time between admissions (p-value < 0.0001; H-value [adjusted for ties] = 186.98). The divergence is explained by the fact that EW admissions vary depending on the weekday and time slot (TS) [47, 48]. The heterogeneity assumption was similarly validated for triage classification time (p-value = < 0.0001; H-value [adjusted for ties] = 14.18) and length of stay in the ICU (p-value < 0.0001; H-value = 12.12). The Kolmogorov-Smirnov (KS) (α = 0.01) and Chi-squared tests (α = 0.01) were used to determine the probability distributions and parameters portraying the random process variables (Table 3).

Table 3.

Results of randomness, homogeneity, and goodness-of-fit test for process variables

Process variables	Levels	Probability expression
Time Between Admissions Randomness (p-value = 0.773; OR = 829; ER = 824.25) Homogeneity (p-value < 0.0001; H-value [adjusted for ties] = 186.98)	Monday – TS1	LOGN(0.0632, 0.156) KS = 0.0714; p-value > 0.75
	Monday – TS2	-0.001 + LOGN(0.0207, 0.0376) KS = 0.0371; p-value > 0.15
	Monday – TS3	-0.001 + LOGN(0.0133, 0.0203) KS = 0.0642; p-value > 0.15
	Tuesday – TS1	LOGN(0.087, 0.215) KS = 0.158; p-value > 0.15
	Tuesday – TS2	-0.001 + LOGN(0.0221, 0.0422) KS = 0.0385; p-value > 0.15
	Tuesday – TS3	-0.001 + LOGN(0.0126, 0.0184) KS = 0.0649; p-value > 0.15
	Wednesday – TS1	WEIB(0.0511, 0.682) KS = 0.104; p-value > 0.15
	Wednesday – TS2	-0.001 + LOGN(0.0252, 0.0455) KS = 0.0421; p-value > 0.15
	Wednesday – TS3	-0.001 + LOGN(0.0134, 0.018) KS = 0.0728; p-value > 0.15
	Thursday – TS1	EXPO(0.0582) KS = 0.0166; p-value > 0.15
	Thursday – TS2	-0.001 + LOGN(0.0297, 0.0642) KS = 0.0619; p-value > 0.15
	Thursday – TS3	-0.001 + LOGN(0.0129, 0.0199) KS = 0.0764; p-value > 0.75
	Friday – TS1	WEIB(0.0486, 0.688) KS = 0.105; p-value > 0.15
	Friday – TS2	-0.001 + LOGN(0.0242, 0.0497) KS = 0.0466; p-value > 0.15
	Friday – TS3	-0.001 + LOGN(0.0137, 0.0221) KS = 0.0933; p-value = 0.0749
	Saturday – TS1	-0.001 + WEIB(0.041, 0.891) KS = 0.089; p-value > 0.15
	Saturday – TS2	-0.001 + LOGN(0.0284, 0.0565) KS = 0.0891; p-value > 0.15
	Saturday – TS3	-0.001 + LOGN(0.0165, 0.0257) KS = 0.0841; p-value > 0.15
	Sunday – TS1	-0.001 + LOGN(0.0528, 0.175) KS = 0.111; p-value > 0.15
	Sunday – TS2	-0.001 + LOGN(0.0253, 0.0484) KS = 0.0651; p-value > 0.15
	Sunday – TS3	-0.001 + LOGN(0.0161, 0.0235) KS = 0.0628; p-value > 0.15
Triage classification time Randomness (p-value = 0.049; OR = 82; ER = 95.48) Homogeneity (p-value < 0.0001; H-value [adjusted for ties] = 14.18)	1 and 2 triage levels	3.5 + 8 * BETA(0.956, 0.739) χ2 = 2.55; p-value > 0.75
	3 to 5 triage levels	POIS(6.69) χ2 = 1.05; p-value > 0.75
Length of stay in EW Randomness (p-value = 0.890; OR = 49; ER = 49.67) Homogeneity (p-value = 0.452; H-value = 7.81)	All admitted patients	UNIF(5, 10)
Length of stay in ICU Randomness (p-value = 0.063; OR = 104; ER = 117.74) Homogeneity (p-value < 0.0001; H-value = 12.12)	Critical patients	EXPO(17.3) χ2 = 4.05; p-value = 0.146
	Patients in moderate condition	EXPO(10.8) χ2 = 6.06; p-value = 0.208

(BETA: Beta; EXPO: Exponential; LOGN: Lognormal; POIS: Poisson; UNIF: Uniform; WEIB: Weibull)

The aforementioned results are embedded into the DT to craft a realistic representation of the EW-ICU pathway. Figure 12 presents the EW submodel contained in the DT. The modelers established a replication time of 15 days/iteration, along with a preceding warm-up period of 100 days. In this context, the model reaches a stable state confirmed by a blocking probability that is close to 0. The synchronization frequency is every minute considering that some respiratory viruses evolve rapidly, provoking a sudden change in patients’ health condition. Likewise, the median end-to-end latency was two minutes, indicating that the DT can reflect a patient’s deterioration two minutes after real occurrence. As the latency is low, decision-makers can suitably monitor patients’ health, obtain early warnings, and then support real-time ICU resource allocations. Thereby, the hospital can minimize the risk of major health complications and the associated mortality probability.

Fig. 12.

EW component of the digital twin

Following this, a pre-sample of 10 runs was carried out to estimate the variability of the main performance indicator (waiting time for an ICU bed). We collected this metric during each run and computed the final number of iterations required to capture the system’s variability. As the required sample size was , there was no need to perform more iterations. The pre-sample covers the system variation (σ = 0.6847 h) and is therefore representative for supporting consequential operational decisions. A 1-sample sign test (α = 0.01) was executed to validate the reliability of the DT with respect to the real system. The outcomes revealed that the simulated model is statistically comparable to the showcased hospital (Ho: η = 68 h || Ha: η ≠ 68 h; p-value = 0.344). The validation was also expanded to queue length (Ho: η = 9.2 patients || Ha: η ≠ 9.2 patients; p-value = 0.754) and patient throughput metrics (Ho: η = 76.2 patients/day || Ha: η ≠ 76.2 patients/day; p-value = 0.34). Therefore, the DT can be employed for process diagnosis and to pretest improvement plans. The current waiting time for an ICU bed ranges from 66.74 to 72.67 h, with a median of 68.86 h. This panorama motivated decision-makers to explore four scenarios:

SC1

Reconfigure an administrative space to create an additional intensive care unit with 3 beds.

SC2

Reconfigure an administrative space to create an additional intensive care unit with 5 beds.

SC3

Reconfigure an administrative space to create an additional intensive care unit with 8 beds.

SC4

Transfer the patients to a partner hospital with an available intensive care unit.

Figure 13; Table 4 provide the results of the Mann-Whitney tests (α = 0.01) performed to compare each proposed operational strategies against the current ICU bed waiting time. In this case, all the strategies were found to be effective for lowering this indicator (p-value < 0.0001; W ≥ 155), ultimately facilitating patient flow management within the EW-ICU pathway. The most effective scenario is SC4, which is projected to slacken this wait between 66.74 and 69.38 h (p-value < 0.001; W = 11,665).

Fig. 13.

Interval plots showing median ICU bed waiting times without intervention and under different operational scenarios

Table 4.

Mann-Whitney results – Comparison of proposed improvement scenarios against current scenario

Comparison	W-value	p-value	Expected difference: Median (95% CI) hours
No intervention Vs SC1	155	< 0.0001	53.31 (48.49; 58.90)
No intervention Vs SC2	155	< 0.0001	60.16 (53.46; 66.08)
No intervention Vs SC3	155	< 0.0001	66.23 (64.70; 69.31)
No intervention Vs SC4	11,665	< 0.001	68.14 (66.74; 69.38)

It is good to note that the DT operated in a closed-loop structure, where patient-level predictions updated the simulation state at 1-minute intervals, and system-related outcomes were fed back into following prediction cycles.

Discussion

The application of predictive models based on artificial intelligence in critical care settings is highly relevant in high-demand clinical scenarios, where early identification of clinical deterioration can directly influence the efficient allocation of resources, patient safety, and health outcomes. In the context of respiratory diseases, recent studies have shown that machine learning algorithms, especially those based on XGBoost, can integrate clinical information and laboratory biomarkers to anticipate the need for intensive care. In this regard, Hernández-Monsalves et al. [49] and Yun et al. [50] demonstrated that the combination of physiological and inflammatory variables substantially improves the prediction of admission to ICU, which supports the relevance and timeliness of the methodological approach adopted in the present study.

From this perspective, the results obtained show that the XGBoost model optimized via Bayesian optimization achieved high predictive performance, with an AUC-ROC of 95.92% (95% CI [93.90% − 97.90%]), along with high levels of accuracy and specificity. This behavior indicates a significant ability to discriminate patients at greater risk of requiring transfer to the ICU within the next 24 h after EW entry, maintaining an appropriate balance between early detection and control of false positives—an aspect critical in contexts of hospital overcrowding. According to the calibration plot, the model demonstrates good agreement across the probability spectrum. The slope is close to 1, indicating that the model has a near-perfect relationship between the increase in predicted ICU need and the increase in actual events. On the other hand, an intercept near 0 denotes that global bias has been largely deleted. Comparable results have been described by Bolourani et al. [51], who observed that XGBoost models outperform logistic regression in the early prediction of respiratory failure, achieving a favorable balance between sensitivity and specificity. Similarly, Sitthiprawiat et al. [1] reported a high discriminative capacity of AI models applied in emergency departments to anticipate both ICU admission and the use of critical resources, which reinforces the external validity of the findings of the present study. This is also advantageous from a financial standpoint, as a balanced model enables better resource allocation. In other words, the model prevents patients who do not require intensive care from being finally transferred to that area, thereby evading unnecessary healthcare costs.

On a different tack, a comparative evaluation was conducted between undersampling and SMOTE-based oversampling to address class imbalance in the ICU request prediction and subsequently prevent evaluation bias. Both approaches demonstrated high discriminatory performance, with comparable ROC area under the curve values, indicating a similar ability to distinguish between classes. As these sampling methods offer a statistical equivalent ROC-AUC, the undersampling choice is justified in this case considering efficiency reasons. First, the model trained with undersampling is simpler which is preferred based on Occam’s Razor principle [52, 53]. Second, the undersampling approach entails shortened training latency, less RAM memory consumption, more efficient production pipelines, and hyperparameter tuning [54]. Also, this technique is less probable to generate data bias in presence of large datasets [55]. Likewise, SMOTE is not advised in this case as it can generate synthetic noise that may lead XGBoost to create overly complex branches to grasp artificial patterns [56, 57]. In this regard, this approach may not reflect the real pathophysiology of a patient needing intensive care.

Also, recent studies have proposed alternative approaches based on advanced heuristic algorithms, such as social network-inspired search methods combined with ensemble models (e.g., SNS-LightGBM), which have demonstrated improvement in global exploration capabilities and model stability in presence of complex and imbalanced data contexts [58].

On a different tack, the threshold trade-off curve (Figure X) exhibits that minor thresholds benefit sensitivity and early detection (early warning), while greater thresholds prioritize specificity and critical resource conservation (ICU scarcity), with the estimated value of 0.43 denoting a balanced operating context. The ICU scarcity scenario is beneficial in case of respiratory epidemics/pandemics where ICU beds and ventilators are near or at capacity [59]. It also favors the rapid influx of critically ill patients to diminish mortality rates and long-term sequelae. In the meantime, the early warning scenario is useful in patients with uncertain respiratory disease trajectories (pneumonia, COVID-19, viral infections) and patients that can worsen rapidly if early hospitalization service is not provided timely.

ANOVA outcomes inform institutional decisions about which laboratory assays and monitoring modalities should be standardized in the emergency workflow to maximize predictive yield. The variable importance analysis showed that laboratory biomarkers, particularly D-dimer and C-reactive protein, play a predominant role in predicting transfer to intensive care (Fig. 14). This result is clinically coherent, since these markers reflect pathophysiological processes central to severe respiratory infections—such as systemic inflammation, endothelial dysfunction, and thrombotic activation—which are widely associated with progressive clinical deterioration [23]. Consistently, Ortiz-Barrios et al. [13] noted that incorporating inflammatory biomarkers into predictive models allows for more precise anticipation of mismatches between the demand for critically ill patients and the available ICU capacity. Inflammatory and thrombotic biomarkers such as D-dimer and C-reactive protein emerge as dominant predictors in the model. Elevated values of these markers are pathophysiologically consistent with systemic inflammation and coagulopathy that predispose to respiratory failure and multi-organ dysfunction in severe viral respiratory disease. Lactate dehydrogenase functions as a surrogate for tissue injury and correlates with the extent of pulmonary involvement, thereby enhancing sensitivity for subclinical progression. Moderate effect sizes were observed for lactate dehydrogenase (d = 0.59), and C-reactive protein (d = 0.53), indicating clinically meaningful differences between ICU-transferred and non-transferred patients.

Fig. 14.

Summary of features with significant influence on ICU transfer probability

Likewise, hemodynamic and respiratory variables, including heart rate and oxygen saturation, showed a relevant contribution, especially when repeated measurements were considered. This finding highlights the importance of capturing the patient’s clinical trajectory, since dynamic changes in vital signs often precede the deterioration that prompts escalation to intensive care. In line with this, Valko and Hauskrecht [60] and Yan et al. [16] emphasized that temporal information provides additional predictive value compared with isolated measurements. Vital signs and their temporal dynamics—notably oxygen saturation and heart rate measured at two time points—provide critical information about trajectory rather than static state; trend information captures early physiological decline that single-point measurements miss. Age retains strong prognostic value as a vulnerability marker, while sex contributes through interaction effects that may reflect differential comorbidity profiles or immune responses.

Finally, sex showed high importance, suggesting that it acts as a risk factor for transfer to the ICU. Likewise, men have 2.1 times higher odds of being in the ICU compared to women. This behavior is consistent with previous reports on high-complexity multivariable models in critical clinical settings, such as those reported by Moore & Bell [23] and Tanaka et al. [24]. Global importance metrics derived from ANOVA inform institutional decisions about which laboratory assays and monitoring modalities should be standardized in the emergency workflow to maximize predictive yield. The combination of local interpretability and global feature ranking facilitates both bedside decision support and strategic data-collection policies. Physics-Informed Neural Networks (PINNs) were also explored by Ma et al. [61] in predicting parameters under extreme environments to enhance the reliability of their model in the context of extreme epidemics. The main results were improved accuracy, generalization, and interpretability, which are highly useful in a healthcare environment.

On a different note, patient-level SHAP explanations convert a model probability into a clinically intelligible attribution of risk drivers, enabling clinicians to identify which features most strongly push a prediction toward high risk and to prioritize targeted interventions. Recent interdisciplinary studies have further demonstrated the applicability of SHAP in complex, high-uncertainty scenarios, such as disaster risk assessment, environmental systems, and mobility forecasting, where it has been used not only for feature attribution but also to evaluate model performance under dynamic and nonlinear conditions [62–65]. These findings reinforce the robustness and versatility of SHAP as an interpretability framework. While these advanced application areas are beyond the scope of this study, they highlight promising directions for expanding explainability in healthcare systems characterized by uncertainty and variability. However, it is good to highlight that SHAP decomposes predictions assuming feature independence, which evidences its limitations in correlated clinical variables [66]. Besides, feature atributions may be unstable across model re-specifications or data perturbations, specially in the presence of multicolinearity, which could affect the consistency of individual-level explanations. At the bedside level, this instability impedes using SHAP explanations in guiding individual clinical decisions or prioritizing variables; instead, it is a transparent metric for reflecting model-specific associations among features and predictions.

Feeding individual risk predictions and empirical predictor distributions into the digital twin enables scenario-based simulation of demand-capacity interactions. The digital twin can quantify the operational impact of interventions such as temporary expansion of ICU capacity, reallocation of staff, activation of intermediate care units, or implementation of intensified monitoring protocols. Predictors that reflect rapid physiological change permit the digital twin to evaluate surveillance policies—such as defined observation intervals and trigger thresholds—and to estimate their effects on waiting times, overflow probability, and expected clinical outcomes. In this way, the digital twin evolves into a prescriptive decision-support platform by linking predictions with operational actions [9, 10, 67]. The integration of ICU transfer probabilities enables the evaluation of interventions under competing objectives, such as reducing waiting times, improving patient outcomes, and controlling resource use [8, 11]. The model’s predictive performance directly shapes these trade-offs: high specificity (94.90%) helps avoid unnecessary ICU admissions (cost efficiency), while adequate sensitivity (81.55%) supports timely identification of high-risk patients (clinical benefit) [1]. Accordingly, the evaluated scenarios represent alternative points within a multi-objective decision space (e.g., cost vs. outcomes), highlighting the need for future incorporation of formal multi-objective optimization approaches [8]. Future iterations of the digital twin should also incorporate ancillary units, such as laboratory and imaging departments, as interconnected service nodes, since diagnostic turnaround times may act as gating conditions for ICU transfer decisions and may propagate delays across the EW–ICU pathway.

The operability of our proposed approach is summarized in Fig. 15. This methodological framework goes beyond simple prediction and extends to decision-making at medical, tactical, and operational levels. However, deployment in other contexts will additionally require minimum data availability and integration into clinical workflows to ensure actionable use. The dataset must include the core features identified in this study along with comorbidity summary, time since symptom onset, oxygen support level, and additional registrations of time-dependent variables. For effective integration into the hospital day-to-day routine, healthcare administrators need to embed the model into the electronic health records and clinical dashboards, granting automatic data ingestion. The model is then run at emergency department triage and 2-hour intervals to detect change points in the patient’s health condition. This deployment should be governed considering that the treating clinician retains the final decision for ICU transfer, while data analytics experts monitor model performance drift (AUC, calibration) and threshold re-assessment. For the transfer-specific workflow, it will be required to checklist the hemodynamic stability and ventilation status of the patient, whilst confirming the internal transportation team availability. Then, ICU receives the transfer request and verify the availability of bed, critical equipment, and supplies.

Fig. 15.

Operability of the optimized AI-digital twin approach

For a prospective validation pathway, a multiphase strategy is proposed comprising: (i) external retrospective evaluation, (ii) silent prospective validation (the model operates without influencing care), (iii) clinical impact study (comparison between standard care Vs. Model-assisted care in terms of ICU bed waiting time and other critical indicators), (iv) safety evaluation (oversight of missed ICU cases, adverse events, and transfer-related complications), (v) equity evaluation (stratification of model performance in terms of age, sex, and socioeconomic status), and (vi) health economic assessment (cost-effective analysis on ICU usage, avoided admissions, and transportation costs).

The conclusions of this research have the power to upgrade clinical performance, reduce mortality rates, and optimize scarce resource usage. In this regard, four scenarios were pretested and concluded as effective if implemented in the wild. The wining solution contemplates transferring patients to a partner hospital with available capacity, emulating the “big hospital” concept and highlighting the need for defining fair collaborative networks. Despite the sharpened waiting time decrease, healthcare protocols for safe transportation and patient reception at the destination institution must be clearly stated to avoid health complications. In this regard, it is fundamental to establish the legal framework governing the cross-hospital patient transfers as they need to be coherent with the applicable healthcare regulations and protocols. The involvement of control bodies, healthcare authorities, and other stakeholders will ensure a feasible application of the collaboration network with efficient flow of information, transfers, and earnings [68, 69]. Another aspect of importance in this line is the capacity-coordination challenges required for executing these collaboration flows. In this respect, the definition of a governance structure led by a steering committee is crucial for addressing the associated patient transfer risks in real-time scenarios. Cross-functional communication protocols along with roles, responsibilities, and service procedures must also be delineated to operationalize this structure effectively [70, 71]. The collaboration also entails the definition of a payment model that satisfies both the source and destination hospitals. This scheme should consider a correlation between the ICU service net incomes and how much resource the transferred patients have used [68, 72]. Additionally, the distribution of patient transfer costs, including transportation between hospitals, patient set-up, communication services, and transfer administration should be previously as part of the collaboration agreement. The transfer eligibility should be also discussed considering two perspectives: patient and destination hospital. Not all the patients admitted in an ICU are qualified for hospital transfer given clinical restrictions (hemodynamically instability, on mechanical ventilation, reception of vasopressors), ethical considerations, logistics difficulties, and transportation risks (restricted monitoring compared to ICU, delays in intervention if complications occur, and environmental stress by movement and vibration). In this case, the transfer benefits should outweigh the risks. On the other side, the destination hospital is eligible whether it can provide a timelier ICU care compared to the source institution, whereas meeting all the transferability conditions. Likewise, model fairness needs to be constantly monitored to avoid access transfer limitations for certain groups given their age, sex, socioeconomic status, and ethnicity. Hospital data culture is also important to this aim as fairness disparities may emerge in patients with poorer data quality. Therefore, data analysts should conduct stratified model performance evaluations, analyze threshold impact by subgroup, and discriminate errors especially during early warning and ICU scarcity scenarios.

Besides, conducting a financial analysis is strongly recommended to ascertain the required investment for each hour or reduced waiting time. This will enable analysts to create a more comprehensive decision-making platform involving both technical and economic aspects.

This study also has certain limitations. First, reliance on two-timepoint measurements enhances temporal resolution but requires standardized timing protocols to avoid measurement-interval heterogeneity. Recording exact inter-measurement intervals and modeling rate-of-change as a continuous covariate would strengthen temporal validity. Furthermore, while time-dependent features were considered in the XGBoost-based model, more robustness is needed to explicitly capture the sequential dependencies in patient flow, which may limit the ability to represent the complex dynamics of the hospital system. External validity is not guaranteed: predictor relevance and model calibration may shift across institutions, epidemic waves, and etiologic agents, so external validation and periodic recalibration are essential for sustained operational utility. Model recalibration is necessary considering the differences in terms of patient profiles, disease presentation, healthcare systems, ICU admission practices, and geographic context. Platt scaling or isotonic regression may be good alternatives to deal with this challenge along with threshold re-optimization. In addition, the current digital twin focuses on the core EW–ICU pathway and does not yet explicitly represent ancillary units such as clinical laboratories and imaging departments. In real hospital operations, these services may become upstream bottlenecks because diagnostic turnaround times, reporting delays, equipment availability, and staffing constraints can postpone clinical decision-making and transfer execution. Future extensions should therefore incorporate these ancillary units as interconnected subsystems within the discrete-event simulation framework, with stochastic service times, queue dynamics, and resource-capacity constraints. Diagnostic completion could then be modeled as a gating condition for selected medical decisions, allowing the digital twin to estimate how delays propagate across departments and affect global patient flow performance. On a different tack, multi-objective optimization models can be employed to enhance scenario analysis by incorporating environmental, social, and economic aspects associated with each proposed scenario. A pending gap in this approach is the interaction with clinicians, which could be tackled through software incorporating both a digital twin and an optimized XGBoost model.

Conclusions

This research evidences the AI-DT power in anticipating the ICU transfer likelihood, diagnosing the operational EW-ICU response, and designing feasible improvement scenarios facilitating patient flow management. This study pinpointed significant predictors that were employed by the XGBoost algorithm for forecasting the need for ICU care in respiratory-affected patients. The model’s prediction capability was upgraded through Bayesian optimization, which allowed for the fine-tuning of model hyperparameters. Going beyond, the SHAP method provided a breakdown of patient risk factors that can be utilized by medical staff and hospital administrators for understanding how certain features push the ICU transfer prediction to specific values of interest. These findings highlight the importance of interpretable AI techniques in extracting relevant insights from complex healthcare datasets, thereby laying the groundwork for improving emergency and intensive care during respiratory seasonal outbreaks.

The predictions derived from the XGBoost model were later incorporated into a digital twin that mimics the healthcare operations taking place from the patient arrival to final process discharges, either absorbing or transitory. These virtual replicas fostered engagement with decision-makers, and facilitated analysis of demand-capacity balance, scarce resource usage, and department-to-department interactions. An additional benefit was the ability to pretest new operational EW-ICU configurations for the detection of potential real-world effects prior to implementation. Our study bridges the gap between data science and public health applications, providing actionable plans that expand the action line of most AI-based implementations.

While this study exhibits a promising approach to addressing the poor response of healthcare systems during respiratory seasonal epidemics, several limitations still remain. First, our research is geographically restricted to Europe, which could affect the generalizability of the insights to other clinical profiles, healthcare structures, and policies. Therefore, we plan to perform external validation on an independent dataset, ideally from a nationwide source across regions such as Africa or South America, to appraise the performance of the ICU need prediction model in different contexts and boost its pertinence across an ample range of medical settings. Besides, the quality of retrospective registrations may restrict our ability to catch all existent data patterns, including social and behavioral aspects, that could fine-tune the model’s prediction accuracy. It is also suggested to explore the use of hybrid models based on deep learning techniques, which have demonstrated greater potential for capturing temporal and nonlinear relationships in complex systems. Frameworks such as Recurrent Neural Networks (RNNs), Long Short-Term Memory (LSTMs), or hybrid models that integrate deep learning with probabilistic approaches could improve the accuracy of predicting sequential patient flow modeling, the temporal evolution of patient flow, and the transition probabilities between states more accurately. The digital twin also holds certain limitations regarding the incorporation of interactions with ancillary units, such as clinical laboratories and imaging departments. Future work will incorporate these services as interconnected subsystems within the discrete-event simulation framework, where diagnostic completion may act as a gating condition for clinical decisions and ICU transfer prioritization. By capturing queue formation, stochastic service times, reporting delays, and resource-capacity constraints in these ancillary units, the digital twin will provide a more comprehensive representation of hospital operations and improve the accuracy of global patient flow predictions between emergency services and the ICU. Deeper analysis on economic cost calculations and medical dispute risk assessments will be performed to provide a more robust decision-making background, especially when implementing cross-hospital transfer plans like SC4. Subsequent research directions will also explore how simulation conclusions change under prediction uncertainty, calibration drift, and different decision thresholds.

In conclusion, this research paves the way for creating a decision-making platform that facilitates real-time, feasible, and effective interventions in the healthcare operability when facing respiratory-disease periods. By capitalizing on the advantages of optimized and interpretable XGBoost models and digital twins, hospital managers and policymakers can better balance operational responses in upstream services (i.e., ICU) where timeliness and accuracy are pivotal for lowering mortality rates and minimizing the risk of long-term sequelae.

Abbreviations

AI

Artificial Intelligence

ANOVA

Analysis of Variance

AUC-ROC

Area Under Curve – Receiver Operating Characteristic

CRP

C-Reactive Protein

DES

Discrete-Event Simulation

DT

Digital Twin

ER

Expected Runs

EW

Emergency Ward

ICU

Intensive Care Unit

KS

Kolomogorov-Smirnov

LDH

Lactate DeHydrogenase

LSTM

Long Short-Term Memory

OR

Observed Runs

RNN

Recurrent Neural Network

SHAP

SHappley Additive exPlanations

TS

Time Slot

XGBoost

eXtreme Gradient Boosting

Author contributions

M.O-B. conceived and supervised the study. M.O-B and S.A-F. developed predictive models and conducted the Bayesian optimization. M.O-B., S.A-F., H.J.C-D., and J.F.T-A. contributed to the methodological framing of the study. M.O-B., A.M-O., I.J.A-G., and M.M.S.B. interpreted and discussed the results. M.O-B., T.A.P.C., L.J.S.C., and O.A.R. were responsible for methodological validation and writing support. All authors reviewed and approved the final manuscript.

Funding

This study received no funding.

Data availability

All the data are available for review upon request.

Declarations

Ethics approval and consent to participate

We confirm that this research was undertaken following the ethical standards established in the 1964 Declaration of Helsinki and its recent adjustments. The project was examined by the Ethical Research Committee of HM Hospitales which approved this study via Agreement #14-12-2021-004 (Access request ID:39). The consent to participate is not applicable.

Consent for publication

Not applicable.

Competing interests

The authors declare no competing interests.