Enhancing Predictive Models to Lower Rehospitalization Risk: Utilizing Historical Medical Records for AI-Driven Interventions

Abstract

Artificial Intelligence (AI) models can predict patient readmission probabilities, aiding discharge decisions and preventing early discharges, which can lead to rehospitalization and related increased costs. EXplainable AI (XAI) methods, like SHAP, can identify important features for the predicted risk. However, the readmission risk and features importance are often considered not informative enough, as they lack in providing suggestions on how to reduce the risk when it is too high to discharge the patient. Our proposed method addresses this gap by using historical medical records of patients with profiles similar to the current case. It generates suggestions for modifying feature values to reduce the AI-predicted rehospitalization risk. By analyzing the historical data distribution of these features, our method infers optimal values and provides targeted suggestions to align current values with these ideals, offering a strategy for situations where discharge risk is deemed too high. In our approach, we explore different definitions of patient similarity to select historical cases and evaluate the effectiveness of the derived suggestions. This is compared with suggestions generated using random historical cases to determine the most effective case selection criteria. Using the MIMIC-III electronic health record dataset, our method demonstrated that two specific sets of historical cases achieved an 80% precision in suggestions, surpassing random case selections by 3%. Ultimately, this method provides recommendations to reduce the readmission risks, serving as a valuable tool in discharge planning. This contributes significantly to reducing premature discharges and the associated costs of rehospitalization.