An Explainable AI-Integrated Diagnostic System for Voice Analysis in Heart Failure Patients

Integrating Explainable Artificial Intelligence to analyse voice characteristics is an essential topic for future research. We explore the utility of tree-based machine learning models, including Random Forest, XGBoost, and LightGBM, in distinguishing between two groups: 100 participants with heart failure and 100 healthy controls. The acoustic features extracted from sustained vowel recordings are used to differentiate between the two groups. The evaluation shows that the Random Forest model performs better, especially with the vowel /i/, achieving Accuracy, Precision, Recall, and F1 score over 0.80. We investigate the interpretability of these models using SHapley Additive exPlanations values, which reveal the essential acoustic features that influence model predictions and provide insights into their clinical relevance. This research highlights the potential of interpretable vocal biomarkers in remote monitoring and diagnosing heart failure.