Development and Validation of a Residual Deep Neural Network for Predicting Vancomycin Trough Concentration Categories in Pediatric Patients

Background Vancomycin is a key antimicrobial for severe gram-positive infections, including methicillin-resistant Staphylococcus aureus, in children. Achieving therapeutic trough concentrations (10–20 µg/mL) is challenging due to pharmacokinetic variability. Conventional dosing and Bayesian models often fail to ensure optimal exposure. Machine learning offers potential improvements, but pediatric-specific applications remain limited. Methods We developed a residual deep neural network to predict categorical vancomycin trough concentrations in hospitalized children using routinely collected clinical data. The model comprised three hidden layers with dropout regularization and residual connections, trained via five-fold group-aware cross-validation to prevent patient-level data leakage. Ensemble averaging of multiple architectures improved generalization. Performance was evaluated using accuracy, F1 scores, area under the receiver operating characteristic curve (AUROC), and area under the precision–recall curve (AUPRC), with interpretability analyzed by SHapley Additive exPlanations (SHAP). Results The dataset included 5,475 therapeutic drug monitoring episodes from 1,753 pediatric patients. The final ensemble achieved 0.860 accuracy, AUROC = 0.984 (95% confidence interval [CI], 0.981–0.988), and AUPRC = 0.950 (95% CI, 0.940–0.960). Most misclassifications occurred between adjacent concentration ranges, indicating physiologically coherent predictions. SHAP analysis highlighted clearance- and dose-related variables—particularly clearance_estimate, estimated glomerular filtration rate, and daily_dose_mg_kg—as major contributors. Conclusions A residual deep neural network with ensemble averaging accurately classified pediatric vancomycin trough categories using standard clinical data. The model demonstrated strong performance, interpretability, and clinical applicability, offering real-time decision support for dose optimization. This framework represents a scalable, data-driven alternative to conventional pharmacokinetic approaches. Future work should validate its multicenter generalizability and clinical integration.