Individualized antibiotic discontinuation support using machine learning: hospital- wide evaluation of gradient boosting and LSTM models

Abstract Background Prolonged antibiotic therapy contributes to antimicrobial resistance and may increase the risk of adverse effects in hospitalized patients. Timely antibiotic discontinuation is an important quality and safety goal; however, identifying the optimal time to stop treatment remains challenging. Machine learning models trained on routinely collected clinical data may provide data-driven decision support to identify patients in whom antibiotic discontinuation could be considered. This study developed and evaluated two machine learning approaches to predict antibiotic discontinuation in hospitalised, non-critically ill patients. Methods A retrospective cohort study was conducted using daily laboratory results and vital signs from hospitalized patients receiving antibiotic therapy at a large academic centre. Two models were developed: an XGBoost model operating on daily patient snapshots enriched with temporal features; and a Long Short-Term Memory (LSTM) neural network processing raw daily clinical sequence. Model performance was evaluated primarily using the Area Under the Precision-Recall Curve (AUPRC), alongside other metrics. Predictions were evaluated using a two-day anticipation window, with predictions considered to be correct when generated on the day of clinician-initiated antibiotic discontinuation or up to two days earlier. Model interpretability was assessed using SHAP (XGBoost) and Integrated Gradients (LSTM). Results The XGBoost model achieved an AUPRC of 0.579, while the LSTM achieved an AUPRC of 0.480 (prevalence-based baseline 0.385) Sensitivity was 41.2% for XGBoost and 19.6% for the LSTM, while precision was 58.5% and 55.5%, respectively. Feature importance analyses showed that infection-related features, such as C-reactive protein and temperature, were among the most influential predictors in both models. Oxygen saturation was also highly influential, alongside data-availability patterns and other clinical variables. Conclusions Both models demonstrated the feasibility of identifying clinical patterns preceding clinician-initiated antibiotic discontinuation from routinely collected clinical data. The two models showed different performances across the evaluated metrics, with XGBoost achieving better performance than the LSTM. By identifying a subset of patients up to two days before clinician-initiated discontinuation, these models could support earlier clinical reassessment of ongoing antibiotic therapy. The models were designed as clinician-in-the-loop decision-support tools rather than autonomous systems. Prospective validation is needed to determine whether model-guided reassessment can safely facilitate earlier antibiotic discontinuation.