Explainable Human–AI Collaboration for Risk-Aware Clinical Decision Support in Transparent and Clinically Meaningful Heart Failure Mortality Prediction
Main Article Content
Abstract
Heart failure remains a major clinical challenge because of its substantial mortality risk and the difficulty of identifying patients who may require early clinical attention. Although machine learning models can support mortality prediction, clinical use requires more than accurate classification: the system must communicate uncertainty and rationale in a form that clinicians can critically evaluate. This study therefore reframes heart failure mortality prediction as a human–AI collaborative decision-support problem in which artificial intelligence provides structured risk evidence while the clinician retains responsibility for contextual interpretation and the final clinical decision. The proposed framework integrates CatBoost machine learning, risk-aware threshold optimization, and SHAP-based explainable artificial intelligence. The Heart Failure Clinical Records Dataset was used, with the temporal variable (time) excluded to reduce information leakage and to restrict prediction to patient characteristics available during clinical assessment. The CatBoost classifier was trained with class-balanced optimization and evaluated using an independent test set and five-fold stratified cross-validation. Youden's J statistic was used to identify a clinically oriented decision threshold, while SHAP generated population-level and patient-level explanations intended to support clinician verification of AI recommendations. The AI component achieved an ROC-AUC of 0.7818 on the independent test set and a mean cross-validation ROC-AUC of 0.7501 ± 0.1327. Threshold optimization identified a cut-off of 0.1844, increasing mortality recall to 89% and detecting 17 of 19 mortality cases. SHAP analysis identified serum creatinine, ejection fraction, and age among the strongest contributors to model output. At the patient level, the framework presents the predicted risk together with factors that increase or decrease that risk, creating an interaction point at which clinicians can accept, question, or override the AI recommendation using additional clinical context. Because this retrospective study did not include a clinician user study, the findings establish the technical feasibility and collaboration-readiness of the proposed interface rather than demonstrating improved human–AI team performance. Prospective evaluation with clinicians and external cohorts is required before real-world deployment.
Article Details

This work is licensed under a Creative Commons Attribution-ShareAlike 4.0 International License.
Authors who publish with International Journal for Applied Information Management agree to the following terms: Authors retain copyright and grant the International Journal for Applied Information Management right of first publication with the work simultaneously licensed under a Creative Commons Attribution License (CC BY-SA 4.0) that allows others to share (copy and redistribute the material in any medium or format) and adapt (remix, transform, and build upon the material) the work for any purpose, even commercially with an acknowledgement of the work's authorship and initial publication in International Journal for Applied Information Management. Authors are able to enter into separate, additional contractual arrangements for the non-exclusive distribution of the journal's published version of the work (e.g., post it to an institutional repository or publish it in a book), with an acknowledgement of its initial publication in International Journal for Applied Information Management. Authors are permitted and encouraged to post their work online (e.g., in institutional repositories or on their website) prior to and during the submission process, as it can lead to productive exchanges, as well as earlier and greater citation of published work (See The Effect of Open Access).