An explainable real-time artificial intelligence -assisted system to reduce intradialytic hypotension

Background Intradialytic hypotension (IDH) is a frequent complication in hemodialysis and is associated with adverse cardiovascular and neurological outcomes. Existing AI prediction models often lack real-world applicability, highlighting the need for clinically relevant, interpretable, and integrated AI systems to support real-time decision-making in hemodialysis care. Methods We analyzed hemodialysis data from 2016 to 2023, incorporating demographics, dialysis settings, and high-frequency vital signs. IDH was defined by a systolic blood pressure drop accompanied by clinical interventions. Machine-learning and time-series models were developed. The best-performing model was integrated into a real-time prediction dashboard with individualized SHapley Additive exPlanations (SHAP) evaluation. This AI system was validated to compare the incidence of IDH with and without AI assistance. Results A total of 1.94 million hemodialysis records from 1351 patients were analyzed. After threshold optimization, CatBoost achieved the highest area under the curve and the best sensitivity–specificity balance. SHAP analysis identified systolic blood pressure and the shock index as key predictors. The system demonstrated a sensitivity of 96.05% and a specificity of 97.16%. The final model was incorporated into a clinical dashboard that provided real-time risk estimation and actionable SHAP-based recommendations. The incidence of IDH decreased significantly from 16.42% to 13.09% after implementation of the AI system (P = 0.03). Conclusion The CatBoost-based AI system can predict IDH with high accuracy while providing individualized, interpretable, and real-time decision support. Application of this explainable AI system was associated with a significant reduction in the incidence of IDH.