Explainable Heart Failure Voice Prediction using Machine Learning Ensembles

Heart failure (HF) detection is one of the challenging health concerns as early diagnosis may reduce mortality rate and improve the quality of life. In this aspect, we propose a novel biomedical voice signal processing framework including multiple voice tasks. The extracted features from multiple voice tasks are used to classify the HF detection using an ensemble of explainable AI (XAI) integrated machine learning classifiers. A key innovation of our methodology is the implementation of a two-stage majority voting strategy to consolidate the predictions of the diverse classifiers across heterogeneous voice tasks. In the first stage, each voice task is independently processed, and predictions from all classifiers are aggregated using standard majority voting; in the second stage, these task-level decisions are integrated via an additional majority voting layer to produce the final HF prediction. This hierarchical voting mechanism is motivated by the need to mitigate bias from any single classifier or voice task, thus enhancing predictive robustness and ensuring that the final decision reflects a consensus derived from multi-task auditory inputs.