AI-DRIVEN RISK STRATIFICATION OF PATIENTS USING ELECTRONIC HEALTH RECORDS

Abstract

Hospital readmissions are a persistent source of preventable harm and inefficiency, motivating the development of accurate and clinically actionable risk prediction tools. This study presents an AI-based risk stratification approach for predicting 30-day hospital readmission using electronic health record (EHR) data, with emphasis on methodological rigor rather than algorithmic novelty. A retrospective cohort of adult patients was analysed using structured EHR variables encompassing demographics, chronic comorbidities, prior healthcare utilisation, laboratory summaries, and medication burden. Features were constructed with strict temporal alignment to minimise information leakage. Model performance was evaluated using a transparent logistic regression baseline and a gradient boosting model. Evaluation incorporated discrimination, calibration, subgroup robustness, and clinical utility analyses. The gradient boosting model demonstrated superior performance, with improved discrimination and more reliable calibration of predicted risks. Visual and quantitative analyses confirmed meaningful separation between readmitted and non-readmitted patients, while subgroup analyses showed stable or improved performance in older adults and across sex categories. Decision curve analysis indicated that model-guided intervention strategies could achieve greater net benefit than non-selective approaches across clinically relevant thresholds. Overall, the results support the potential of carefully evaluated AI-driven risk stratification to inform targeted discharge planning, while highlighting the need for external validation prior to real-world deployment.