Predicting Unplanned Readmission Risk in Patients With Cirrhosis: Complication-Aware Dynamic Classifier Selection Approach.

Background: Cirrhosis is a leading cause of noncancer deaths in gastrointestinal diseases, resulting in high hospitalization and readmission rates. Early identification of high-risk patients is vital for proactive interventions and improving health care outcomes. However, the quality and integrity of real-world electronic health records (EHRs) limit their utility in developing risk assessment tools. Objective: Despite the widespread application of classical and ensemble machine learning for EHR-based predictive tasks, the diversity of health conditions among patients and the inherent limitations of the data, such as incompleteness, sparsity, and temporal dynamics, have not been fully addressed. To tackle those challenges, we explored a framework that characterizes patient subgroups and adaptively selects optimal predictive models for each patient on the fly to enable individualized decision support. Methods: The proposed framework uniquely addresses patient heterogeneity by aligning diverse subgroups with dynamically selected classifiers. First, patient subgroups are generated and characterized using rules indicating medical diagnosis patterns. Next, a meta-learning framework trains a meta-classifier for optimal dynamic model selection, which identifies suitable models for individual patients. Notably, we incorporated a tailored region of competence to refine model selection, specifically accounting for cirrhosis complications. This approach not only enhances predictive performance but also elucidates why individualized predictions are better supported by selected classifiers trained on specific data subsets. Results: The proposed framework was evaluated for predicting 14-day and 30-day readmission in patients with cirrhosis using multicenter data obtained from 6 hospitals. The final dataset comprised 3307 patients with at least 2 admission records, along with a range of factors including demographic information, complications, and laboratory test results. The proposed framework achieved an average AUC (area under the curve) improvement of 5% and 4% compared to the best baseline models, respectively. Conclusions: By leveraging the expertise of the most competent classifiers for each patient subgroup, our approach enables interpretable training and dynamic selection of heterogeneous predictive models. This advancement not only improves prediction accuracy but also highlights its considerable potential for clinical applications, facilitating the alignment of diverse patient subgroups with tailored decision-support algorithms.