AI-driven voice analysis for predicting heart failure severity: a prospective study

초록

Background: Hypervolemia in heart failure (HF) leads to edema in key speech-producing organs, including the lungs, larynx, and articulatory structures. Specifically, pulmonary edema affects voice resonance and acoustic characteristics. Purposes: This study aims to develop and validate a deep-learning model that monitors HF severity and predicts acute decompensated states by analyzing longitudinal voice data collected during the treatment process. Methods: This multicenter prospective clinical study involves 124 participants (92 HF patients admitted for acute exacerbation and 32 non-HF controls). Voice recordings (vowel prolongation, words, sentences, and free speech) are collected at five time points for the HF group (admission, discharge, and weeks 1, 5, and 13 post-discharge) and two time points for the control group. Simultaneously, clinical data including NT-proBNP levels, NYHA functional class, dyspnea scale scores, and echocardiographic findings will be gathered. An AI model will be trained on extracted acoustic features and clinical variables to classify HF severity into four categories: normal, mild, moderate, and severe. Results: The study expects to identify specific vocal biomarkers that significantly correlate with the degree of pulmonary congestion and clinical stability. It is hypothesized that deep-learning models trained on these vocal features will accurately distinguish between compensated and decompensated HF states, mirroring changes in NT-proBNP levels and clinical findings. Conclusions: Voice analysis holds potential as a non-invasive, cost-effective, and accessible tool for monitoring heart failure severity. This study will provide the foundational evidence for AI-driven digital health tools aimed at early detection of HF aggravation and personalized symptom management.