Explainable machine learning framework for congestive heart failure detection using GLCM-based HRV features and Bayesian network analysis

Background Congestive heart failure (CHF) is one of the main causes of cardiovascular morbidity and mortality. While conventional machine learning (ML) models demonstrate acceptable classification results, their operation is typically based on a black-box approach which does not allow to obtain an interpretable insight into the physiological mechanism of heart rate variability (HRV). Objective This paper suggests an approach which combines Gray-Level Co-occurrence Matrix (GLCM)-based texture analysis, Bayesian-optimized ML and Bayesian network inference to increase the accuracy and clinical interpretability of CHF prediction and provide a deeper understanding of physiological mechanisms. Methods Two-dimensional representation of HRV data taken from the PhysioNet database was used for extracting texture-based biomarkers using GLCM. Most discriminatory biomarkers were selected by the Kruskal-Wallis test and ML models were optimized by Bayesian approach. Performance of models was estimated by traditional classification metrics and Bayesian network analysis with posterior probabilities and sensitivity analysis of nonlinear dependencies between HRV texture-based features was performed. Results Best performance was achieved by Bayesian-optimized Quadratic Gaussian SVM, with 93.10% accuracy, 90.91% sensitivity, 94.44% specificity and AUC equal to 0.9697. The Bayesian network analysis revealed Dissimilarity, Cluster Prominence and Cluster Shade as the most important HRV texture biomarkers related to CHF. Conclusion The novel XAI architecture is based on Bayesian optimization of ML along with Bayesian probability theory, which helps deliver precise and clinically relevant CHF diagnosis for early disease detection and personalized risk stratification.