ERAF-Net: An Explainable Reliability-Aware Adaptive Fusion Framework for Trustworthy Cardiovascular Risk Assessment

Abstract

Cardiovascular diseases continue to be one of the most prevalent causes of mortality across the globe, requiring the design of advanced clinical decision support systems that should not only be accurate but also reliable, interpretable, and trustworthy. Although modern machine learning approaches show promising results in predicting cardiovascular risk, their opaque nature and lack of reliability measurement make their clinical application difficult. To overcome these problems, we propose ERAF-Net: an Explainable Reliability-Aware Adaptive Fusion Framework for Trustworthy Cardiovascular Risk Assessment. This approach combines heterogeneous predictors with the use of a novel Reliability-Aware Meta Fusion Network (RAMFN), where fusion weights are estimated based on a Clinical Reliability Evaluator (CRE). This CRE algorithm estimates the level of trustworthiness of the prediction made by a particular model by taking into account model confidence, significance of explanations, and data availability. An Explainability-Aware Confidence Calibration (EACC) method is suggested to calibrate confidence of predictions in line with explanation consistency. The extensive experiments carried out using public cardiovascular data sets show that the ERAF-Net is highly reliable in terms of predictions compared to the traditional machine learning methods by demonstrating improved accuracy estimation, explainability, and credibility. The experimental results indicate the ability of the ERAF-Net model to be utilized in providing clinical decision support that is transparent and credible.