Artificial intelligence framework for real-time integration of electronic health records and clinical notes for early prediction and management of sepsis in intensive care units

Abstract

Sepsis and septic shock are medical emergencies with associated high mortality rates that require rapid identification and timely action. Recent developments in artificial intelligence (AI) may enable technologies capable of early detection of sepsis and septic shock. This study describes a hybrid AI system that combines deep learning (DL), Bidirectional Long Short-Term Memory (BiLSTM), gradient boosting (XGBoost), and reinforcement learning (Deep Q-Network [DQN]), presented as a dynamic management system capable of real-time prediction as well as management of sepsis and septic shock. In total, we trained the hybrid AI system on over 50,000 ICU admissions from the MIMIC-IV dataset. Our model predicted the onset of sepsis approximately 5.8 h before clinical recognition, with an area under the receiver operating characteristic (AUROC) of 0.91, precision of 0.85, recall of 0.81, and an F1-score of 0.73. Incorporation of structured electronic health record (EHR) data and integration of unstructured clinical notes using ClinicalBERT improved early sepsis detection performance. Retrospective simulated evaluation of the reinforcement learning module suggested a potential 7.2% improvement in estimated survival outcomes compared with standard care policies; however, prospective clinical validation remains necessary. Previous studies have explored how hybrid AI can support real-time, interpretable, and clinically actionable decision support systems in Intensive Care Unit (ICU) settings. The proposed method outperforms existing methods and demonstrates potential to support clinical decision-making in a simulated environment.