Voice features and deep learning models for identifying acute decompensated heart failure

Background Acute decompensated heart failure (ADHF) is a state of systemic congestion which frequently requires hospitalization. Early and noninvasive detection of decompensation remains challenging. This study evaluated whether voice characteristics differ between patients hospitalized with ADHF and after recovery to examine the potential of voice as a non-invasive biomarker for ADHF. Methods Patients requiring hospital admission for ADHF were prospectively enrolled. Patients with respiratory infection, sepsis, lung or vocal cord disease, acute coronary syndrome, or serum creatinine>3.0mg/dL were excluded. Voice was recorded at two time points: at admission for ADHF and at discharge after recovery. Low-level audio features were extracted, and Mel-spectrograms derived from the voice waveforms were used as the input feature for deep learning-based classification models. Model performance was primarily evaluated at the patient-state level using area under the receiver operating characteristic curve (AUC) with 95% confidence intervals estimated by patient-level bootstrap resampling. Results A total of 100 patients were included (mean age 70.5 years; 47% women) and randomly divided into a training group (N=88) and a test group (N=12). In the analysis with low-level audio features, harmonics-to-noise ratio and Shimmer demonstrated discriminative potential between decompensated and recovered states. In patient-state-level analysis, the best-performing model, ConvNeXt-large, achieved a mean AUC of 0.83 (95% CI 0.70-0.95). Segment-level analysis demonstrated higher discriminative performance (AUC up to 0.87), suggesting the presence of discriminative acoustic features. Conclusions Voice characteristics differed between admission and discharge states in patients hospitalized for ADHF, and voice analysis using deep learning models was able to distinguish acute decompensated and recovered states in patients with ADHF. Our results suggest that voice analysis may provide a novel and non-invasive approach for detection of ADHF.