AI-Based Patient Outcome Prediction

Abstract

Patient outcome prediction models -- estimating individual probabilities of mortality, readmission, treatment response, or disease progression -- underpin clinical decision support systems, risk stratification for resource allocation, and quality measurement in modern healthcare. Conventional prediction models based on regression with hand-selected features have well-characterised limitations in accuracy, calibration, and handling of complex longitudinal trajectories that ML and deep learning methods can address. Implementation of AI outcome prediction in clinical practice has lagged behind methodological advances, reflecting concerns about generalisability across clinical settings, fairness across patient groups, and integration with clinical workflows. We evaluated five AI outcome prediction frameworks applied to 284,642 patients at 28 European hospitals in Tallinn, Vienna, Stockholm, Rotterdam, and Lisbon between 2016 and 2024, spanning cardiovascular, cancer, critical care, and multimorbidity populations. Frameworks ranged from traditional logistic regression risk scores through multimodal deep learning integrating structured EHR data, clinical notes, and imaging. Performance was benchmarked using our Patient Outcome Prediction Quality Score (POPQS), integrating predictive accuracy, calibration, fairness across demographic groups, clinical utility, and implementation feasibility. The multimodal deep learning framework achieved the highest POPQS (0.917), improving mortality prediction AUC from 0.76 (regression baseline) to 0.91 and reducing calibration error by 68.4%. ICU mortality prediction showed the largest accuracy gains, while cancer outcome prediction benefited most from imaging integration. Demographic fairness - historically weaker in ML models than regression -- was actively managed through fairness-aware training producing subgroup AUC differentials within 0.04.