This retrospective study included two pediatric cohorts. The internal cohort comprised 1,697 patients under 18 years of age admitted to Guangzhou Women and Children's Medical Center (GWCMC) from February 2016 to July 2018. Pediatric sepsis was identified based on modified Sepsis-III criteria (18), with adjustments for age and clinical presentation. The Sepsis-3 consensus definition provided a robust framework for identifying sepsis and septic shock, which we adapted for pediatric use in this study (19). Electronic health record (EHR) data were extracted, encompassing vital signs, laboratory results, medications, and nursing assessments within the 12-h window preceding a sepsis diagnosis. Cases with incomplete records, neonates, or acute upper respiratory infections were excluded.

The external validation cohort was derived from the MIMIC-III database, consisting of 827 pediatric ICU encounters fulfilling Sepsis-III criteria (17). Variable definitions and coding structures were harmonized across cohorts to ensure consistency. Figure 2 illustrate variable frequency and inter-variable correlations, respectively.

Variables with >20% missingness were excluded from model development. As summarized in Table 1 (Additional File 1), a total of 28 variables were retained, including core vital signs (HR, RR, SpO₂, T), key laboratory biomarkers (WBC, creatinine, lactate, pH), therapeutic indicators (antibiotics, glucocorticoids, vasopressors), and fluid balance measures. Sixteen variables (e.g., bilirubin, troponin, IL-6, PaCO₂) exceeded the 20% missingness threshold (25%–50%) and were therefore excluded. This structured selection ensured that retained features were both clinically relevant and statistically reliable.

We first analyzed the temporal distribution of dynamic clinical variables to understand sampling irregularity across patients. As shown in Figure 3, time intervals for key laboratory and physiological variables such as creatinine (Cr), glucocorticoids (GCS), and pH were highly heterogeneous, ranging from a few hours to over 96 h across individuals. This motivated the adoption of a histogram-based strategy for imputation window selection. To determine the optimal number of reference windows (N), we analyzed the temporal correlation structure of key physiological variables. As illustrated in Figure 4, variables with strong autocorrelation supported extended interpolation windows.

To address missingness, we implemented a dual-step imputation method combining correlation-weighted continuous time windowed histogram (CTWH) estimation with multivariate Gaussian processes (MGP). This approach preserved intra- and inter-variable patterns.

To further illustrate the irregularity of temporal sampling across different clinical variables, we plotted the hourly measurement frequencies of 24 representative features (Additional File 2). Vital signs such as HR and SBP were densely sampled, while laboratory indicators including lactate and creatinine showed sparse and heterogeneous recording patterns. This observation further justified our use of correlation-enhanced CTWH + MGP interpolation to robustly impute time series across variable horizons.

Model development consisted of two stages. In the first stage, a gated recurrent unit (GRU)-based recurrent neural network (RNN) was trained to encode dynamic time-series features. The RNN included two GRU layers (64 hidden units each), optimized using Adam with early stopping.

In the second stage, an XGBoost classifier was trained on either the raw interpolated features or the RNN-derived embeddings. Hyperparameters were selected via 5-fold cross-validation, with early stopping at 10 rounds. A total of 13 independent models were trained, each corresponding to a specific hour (T = 0 to T = 12) before diagnosis. Figures 5, 6 display performance comparisons and statistical evaluations.

The hybrid modeling architecture is further illustrated in Additional File 3, where temporal embeddings extracted by the RNN are combined with static features before being passed into the XGBoost classifier. SHAP analysis was then applied to quantify feature contributions.

Performance metrics included AUROC, AUPRC, sensitivity, specificity, Brier score, and Youden's index. The XGBoost model achieved its highest AUROC (0.915) at T = 0 h, while the RNN demonstrated stability at earlier horizons, peaking at 0.902 at T = 8 h. Results are detailed in Table 2.

Model explainability was assessed via SHAP (Shapley additive explanations) (20). Top-ranked predictors included lactate, pH, white blood cell count, fluid balance, and vasopressor usage (Figure 7). External validation on the MIMIC-III dataset yielded comparable performance trends (Figure 8). While Figure 7 highlights the top contributors to model output as determined by SHAP values, a comprehensive statistical comparison of all candidates features between sepsis and non-sepsis groups is provided in Additional File 4. Extended embedding-level feature contributions and temporal heatmaps are further detailed in Additional File 5.

The final system was configured as a real-time clinical decision support tool. Prediction scores were stratified into three alert tiers: low (0.5≤ score <0.6), medium (0.6≤ score <0.8), and high (≥0.8). Each alert level was linked to specific clinical response protocols.

Retrospective validation demonstrated strong agreement between predicted alerts and physician-confirmed sepsis diagnoses (Cohen's κ = 0.82). Notably, in high-risk cases, alerts preceded treatment initiation by up to 10.41 h, indicating meaningful potential for anticipatory intervention (21).

To facilitate reproducibility, we provide pseudocode describing the complete pipeline, including data preprocessing, feature engineering, model training, validation, and SHAP-based interpretability (Additional File 6).

A total of 1,697 pediatric patients were included in the internal cohort from Guangzhou Women and Children's Medical Center (GWCMC), among whom 444 met the Sepsis-III diagnostic criteria during hospitalization, accounting for 26.2% of the cohort. The median age was 1.88 years [interquartile range (IQR), 0.3–4.82], and the proportion of male patients was significantly higher in the sepsis group than in the non-sepsis group (P = 0.035, χ² test). Baseline demographic and clinical features of the internal cohort are summarized in Table 3. The external validation cohort comprised 827 pediatric ICU patients extracted from the MIMIC-III database, screened using pediatric-adjusted Sepsis-III criteria and confirmed independently by two pediatric intensivists. Static and dynamic variables were harmonized across both datasets. Beyond the sepsis-confirmed validation cohort (n = 827), the broader MIMIC-III pediatric ICU dataset (n = 3,308) exhibited a wide spectrum of discharge diagnoses, including pneumonia (n = 797), fever (n = 479), hypotension (n = 326), and sepsis (n = 66). The full distribution of diagnoses is provided in Additional File 7, highlighting the heterogeneity of the external dataset and supporting the generalizability of our model. The distribution of available laboratory and physiological indicators within 72 h is summarized in Additional File 8, highlighting the heterogeneity and sparsity of EHR data inputs.

We first compared the performance of different data imputation strategies. The combination of continuous time-window histogram (CTWH) and multivariate Gaussian process (MGP) yielded the highest accuracy in imputing sparse variables and led to improved downstream model performance. Notably, in early prediction windows (≥6 h before diagnosis), CTWH + MGP significantly outperformed single-method interpolations. At T = 0 h, the CTWH + MGP-based model achieved an AUROC of 0.915, which was significantly higher than the baseline method (AUROC = 0.882; P < 0.01). The superiority of the CTWH + MGP strategy over baseline interpolation methods was confirmed via significant improvements in downstream model performance, as illustrated in Figure 8. Detailed evaluation metrics of the MGP-based model across prediction horizons are presented in Table 4.

In addition to accuracy, computational efficiency was evaluated across interpolation methods (Additional File 9). CTWH + MGP achieved comparable AUROC to MGP alone (0.829 vs. 0.832, p > 0.05), while reducing average input dimensionality by two-thirds and training time by approximately 68%. This balance of accuracy and efficiency further supports the feasibility of CTWH + MGP for real-time clinical deployment.

We assessed the hour-by-hour performance of both models over a 12-h forecasting window prior to sepsis diagnosis (T = –12 h to T = 0 h). The XGBoost classifier, trained on features interpolated using the CTWH + MGP method, achieved its peak discriminative performance at T = 0 h, with an AUROC of 0.915, sensitivity of 0.88, and specificity of 0.84. In contrast, the RNN model showed slightly lower AUROC values in short-term prediction windows but demonstrated more stable performance at longer horizons, maintaining AUROC values above 0.80 beyond T = –10 h.

Figure 9 presents the AUROC trajectories and corresponding 95% bootstrapped confidence intervals across time points, highlighting the trade-off between short-term accuracy and long-term robustness. Additionally, comparative analysis of model architectures and interpolation strategies further supports the generalizability of the RNN + CTWH + MGP combination across both internal and external datasets. Summary statistics for AUROC, AUPRC, sensitivity, specificity, accuracy, and Youden's Index at each time point are detailed in Table 5. Extended evaluation metrics for the XGBoost model across all prediction windows (0–12 h) are provided in Additional File 10, and the corresponding full RNN metrics are available in the Supplementary Materials. Together, these detailed results confirm the temporal dynamics of model performance, with extended summary comparisons available in Additional File 7.

As illustrated in Additional File 11, although all models achieved comparable AUROC near the time of diagnosis (T = 0–2 h), the RNN + PCC + MGP model maintained significantly higher predictive stability across longer horizons (>6 h). In contrast, XGBoost-based models demonstrated a steep decline in AUROC, highlighting the advantage of temporal modeling for long-range prediction.

To further evaluate temporal model architectures, we compared RNN with more advanced recurrent variants (LSTM and GRU). As shown in Additional File 12 and visualized in Additional File 13, all three achieved nearly identical AUROC values across prediction horizons, with only marginal improvements (<0.01 AUROC) for LSTM and GRU compared with RNN. Given the negligible performance difference, the simpler RNN was adopted for the main analysis due to its computational efficiency.

To further evaluate generalizability, we stratified model performance by age groups in both internal and external cohorts (Additional File 14). In the internal cohort, younger patients (<1 year) consistently showed the highest AUROC values (0.93 at T = 0 h, remaining >0.79 at T = 12 h), whereas older children (>12 years) demonstrated relatively lower discrimination. External validation with MIMIC-III revealed more heterogeneous patterns across age strata, with peak AUROC values observed in the 7–12 year group (0.875 at T = 0 h; 0.860 at T = 6 h). Importantly, differences in AUROC between internal and external datasets remained modest (<0.05 across all horizons), supporting the robustness and transportability of the proposed framework across age subgroups.

Feature contribution was assessed using SHAP (Shapley additive explanations). Across the 12-h prediction window, lactate, pH, white blood cell count (WBC), fluid balance, and vasopressor administration consistently emerged as the most influential predictors of impending sepsis. In longer horizons, dynamic physiological indicators such as respiratory rate and cumulative fluid intake gained relative importance. Notably, all top predictors identified by SHAP analysis (Figure 7) originated from the retained 28 features, reinforcing the validity of our variable selection strategy (see Additional File 1). Beyond overall ranking, we further examined the temporal dynamics of feature contributions to better capture evolving clinical signals.

As illustrated in Additional File 15, feature contributions demonstrated marked variation across time windows. Lactate peaked as the dominant predictor at T = –4 h (SHAP = 0.84) and remained highly influential at T = –2 h and T = 0 h, while heart rate importance increased sharply closer to diagnosis (T = –2 h and T = 0 h). In contrast, systolic blood pressure (SBP) and WBC showed moderate but fluctuating contributions, and respiratory rate and temperature remained relatively minor predictors. These temporal patterns highlight lactate and heart rate as the most reliable early-warning biomarkers in the hours preceding sepsis onset.

In addition, the relative contribution of features varied across prediction horizons. As summarized in Additional File 16, short-term predictions (0–2 h before onset) were more strongly influenced by therapeutic interventions (e.g., glucocorticoid use) and acute biomarkers (e.g., lactate, creatinine), whereas long-term horizons (2–12 h) were dominated by sustained metabolic indicators such as lactate and creatinine. Together, these findings underscore the dynamic and multi-faceted nature of sepsis progression, where both acute hemodynamic changes and longer-term metabolic disturbances contribute to the discriminative ability of the model.

To further clarify the interpretability pipeline, we provide an additional schematic (Additional File 17), illustrating how temporal embeddings from the RNN were combined with static features for XGBoost classification, followed by SHAP/LIME analysis to produce clinically actionable insights. Moreover, embedding-level contributions are detailed in Additional File 5, where the top 15 latent temporal embeddings are ranked by SHAP importance, and a temporal heatmap illustrates how clinical variables (e.g., lactate, creatinine, heart rate) dynamically vary in predictive weight across the 12-h forecasting horizon. These results underscore the complementary role of latent embeddings and raw clinical features in shaping the discriminative power of the hybrid RNN–XGBoost framework.

In the MIMIC-III external validation cohort, the XGBoost model retained strong performance, achieving an AUROC of 0.905 at T = 0 h, while the RNN model exhibited stable prediction (AUROC = 0.88 at T = 8 h). These findings were consistent with internal results. Figure 8 compares the temporal evolution of AUROC and AUPRC across both cohorts.

Additionally, Table 2 summarizes the cross-dataset AUROC performance of both models at various forecast windows. Table 6 presents the deviations in indicator frequency and variable importance between the internal (Guangzhou Women and Children's Medical Center) and external (MIMIC-III) cohorts across cumulative time windows (2 h, 6 h, 8 h). Notably, features like temperature (T), white blood cell count (WBC), C-reactive protein (CRP), and direct bilirubin (DBIL) showed low relative weights in both datasets, suggesting limited predictive influence.

Building on these validation results, we next evaluated the potential bedside impact through retrospective real-time simulation on historical EHR sequences. High-risk patients (predicted probability ≥0.80) were identified a median of 6.2 h prior to physician-confirmed sepsis recognition. Model alerts demonstrated strong concordance with clinical diagnoses (Cohen's κ = 0.82).

Importantly, patients who triggered early alerts exhibited lower rates of delayed ICU transfer and reduced incidence of respiratory failure, underscoring the clinical relevance of timely detection. These findings are summarized in Table 7, which details stratified outcome improvements associated with early-warning interventions.

To translate these results into practice, we designed a tiered alert system to guide clinical escalation (Additional File 18). Predicted probabilities were mapped to three levels of response: •

Tier 1 (P > 0.65): Nurse notification to increase bedside vigilance.