AI-Based Mortality Risk Stratification

Abstract

AI-based mortality risk stratification models identify patients at elevated risk of death across diverse clinical contexts: hospital inpatients at risk of unexpected deterioration, chronic disease patients at risk of short-term mortality requiring goals-of-care discussions, and ambulatory patients at risk of adverse cardiovascular or cancer outcomes requiring preventive intervention. The clinical use cases for mortality risk stratification vary substantially in appropriate deployment pattern, required accuracy, ethical considerations, and action implications. Poorly designed mortality risk stratification systems can cause substantial harm: inappropriate resource allocation, premature goals-of-care conversations, stigmatisation of patients flagged as high-risk, and exacerbation of health disparities when models perform differently across demographic subgroups. We evaluated five AI mortality risk stratification frameworks applied to 142 deployments across 22 European hospitals and clinical networks between 2019 and 2024, covering inpatient deterioration, chronic disease prognostic, and ambulatory cardiovascular/cancer risk applications. Frameworks ranged from simple score deployment through integrated AI risk stratification with ethical governance, appropriate intervention pathways, and equity monitoring. Performance was benchmarked using our Mortality Risk Stratification Effectiveness Score (MRSES), integrating discrimination, calibration, clinical action appropriateness, equity across subgroups, and ethical acceptability. The integrated governed framework achieved the highest MRSES (0.912), with discrimination AUC 0.88, appropriate clinical action in 82.4% of flagged cases, and minimal equity gaps across demographic subgroups. Inpatient deterioration models showed the strongest technical performance while ambulatory prognostic applications showed the largest ethical complexity.