Abstract 11315: Clinical Utility of Machine Learning-Derived Vocal-Biomarker in the Management of Heart Failure

Abstract

Introduction: Previous studies have suggested that early therapeutic intervention could prevent heart failure (HF) worsening, while early detection of signs and symptoms of cardiac decompensation remains challenging. For the first time, this study focused on voice symptoms and aimed to explore its potential utility in the management of heart failure. Methods: This single-center, prospective, observational study included 38 patients admitted due to worsening HF. A total of 726 voice data were recorded during the therapeutic time course at the hospital. Changes in clinical symptoms (New York Heart Association [NYHA])) evaluated by an experienced cardiologist and B-type natriuretic peptide (BNP, pg/ml) during the time course were also evaluated. Here we tested if the acoustic features of the patient voice could predict the NYHA class and BNP levels, using the machine learning (ML) models trained with a gradient boosting algorithm. A 5-fold cross-validation was applied to evaluate the model performance. Results: There were significant correlations of some acoustic features with clinical symptoms and BNP levels (data not shown). We created the prototype of the ML model on the acoustic features extracted from the data set. Receiver operating characteristic (ROC) analysis revealed a good diagnostic accuracy of our ML model to predict NYHA class ≥2 (sensitivity 0.78, specificity 0.75, area under the curve (AUC) 0.79, and accuracy 0.77). There was also a significant correlation between voice-derived BNP and actual BNP levels (Figure). AUC of ROC analysis for predicting NYHA class ≥2 from BNP was 0.688. Conclusions: The prototype of the ML model was related to clinical symptoms of HF decompensation and BNP levels, demonstrating greater diagnostic accuracy for NYHA prediction than actual BNP levels. Further studies are warranted to investigate the possible contributions of our ML model to the early detection of worsening HF as ML-derived vocal-biomarkers.