INTERPRETABLE MACHINE LEARNING FOR EARLY MORTALITY PREDICTION IN ICU PATIENTS USING DYNAMIC VITAL SIGNS, LABORATORY TRENDS, AND TREATMENT PATTERNS

Abstract

Cardiac mortality prediction is very important in intensive care unit (ICU) patients to take timely clinical interventions, planning and resource allocation. In the paper entitled: Mortali: Interpretable Machine Learning for Early Mortality Prediction in ICU Patients Using Dynamic Vital Signs, Laboratory Trends, and Treatment Patterns, an interpretable machine learning (iML) approach to early mortality prediction for patients in the Intensive Care Unit (ICU) is presented, using continuously collected vital signs, laboratory trends, and treatment-pattern. Unlike traditional static risk scoring systems, the proposed approach is dynamic and can reflect changes over time in a patient's condition, based on dynamic vital signs, lab trends, and dynamic patterns of clinical intervention. Two of the models were designed and tested to see their reliability in predicting, to ultimately choose the best model. The interpretability of the model was discussed using the techniques of Explainable Artificial Intelligence (XAI) to determine the most influential clinical factors contributing to the risk of death. The findings suggest that real-time patient monitoring and machine learning with interpretable models could be valuable tools in improving risk stratification and clinical decision-making in critical care settings. The proposed framework can help clinicians in the ICU to detect high-risk patients earlier and optimise patient outcomes by timely intervention, as it offers clear and transparent predictions.