Accepted for/Published in: JMIR Medical Informatics
Date Submitted: Jun 24, 2024
Date Accepted: Jul 15, 2025
(closed for review but you can still tweet)
Warning: This is an author submission that is not peer-reviewed or edited. Preprints - unless they show as "accepted" - should not be relied on to guide clinical practice or health-related behavior and should not be reported in news media as established information.
Complications-aware Dynamical Classifier Selection for Unplanned Readmission Risk Prediction in Patients with Cirrhosis
ABSTRACT
Background:
Cirrhosis is the leading cause of non-cancer deaths among gastrointestinal diseases leading to substantial rates of hospitalization and readmission.
Objective:
Early identification of high-risk patients enables proactive interventions to improve the healthcare outcome. However, this task is challenging due to the inherent limitations of real-world data such as incompleteness, sparsity and temporal dynamics, and the diversity of health conditions among patients. To tackle these challenges, we developed a framework for EHRs-based predictive tasks to enable subgroup-specific decision making by incorporating dynamic classifiers selection.
Methods:
The proposed framework focuses on developing and tailoring classifiers for diverse patient subgroups by dynamic classifier selection, which is enhanced by capturing the patterns of complications and comorbidities. In comparison to existing studies, this study extracts interpretable rules representing complication and comorbidity combinations to enhance the generation of the pool of classifiers, enabling the characterization of patient subgroups and providing insights into why individualized predictions are better supported by classifiers trained on specific subsets of data. Futhermore, we incorporated new meta-features utilizing medical diagnosis-based regions of competence to support the dynamic selection of classifiers within the META-DES framework.
Results:
The proposed framework was evaluated for the prediction of 14-day and 30-day readmission in patients with cirrhosis using multi-center data obtained from six hospitals. The final dataset used for evaluation included 3,307 patients with at least two admission records and various factors such as demographic information, complications and laboratory test results. The presented framework demonstrated superior predictive performance compared to the baseline models used in this study.
Conclusions:
The proposed framework can improve the performance by leveraging the expertise of the most competent classifiers for each subgroup, allow for training and dynamic selection of heterogeneous classifiers in a highly interpretable manner, and emphasize its potential value in clinical settings for aligning the diversity of patient subgroups with the dynamic selection of classifiers.
Citation
