The voice of decompensation: an explainable machine learning approach to assessing heart failure severity.

AIMS: Heart failure (HF) is characterized by high morbidity and frequent hospital readmissions, highlighting the need for scalable out-of-hospital monitoring to support post-discharge nursing care. We aimed to (i) develop a HF-tailored, user-friendly voice task set suitable for remote monitoring; (ii) identify HF-related acoustic markers and build explainable voice-based models; and (iii) evaluate the accuracy of smartphone recordings for clinical monitoring. METHODS AND RESULTS: In this multicentre observational study, patients hospitalized with acute heart failure (AHF) were recruited at four centres. Voice recordings were obtained with professional devices and common smartphones (Apple, Huawei, Vivo). Eleven speech tasks were evaluated, and 11 acoustic feature categories were extracted [e.g. Mel frequency cepstrum coefficients (MFCCs), chroma, spectral, glottal features]. Two XGBoost models were trained to classify clinical status from admission to discharge and from mid-hospitalization to discharge; models were interpreted with Shapley Additive Explanations (SHAP). Four tasks were selected for the final model. The admission-to-discharge model achieved 0.76 accuracy, 0.86 sensitivity, 0.65 specificity, and an area under the ROC curve (AUC) of 0.77. The mid-hospitalization-to-discharge model showed 0.81 accuracy, 0.89 sensitivity, 0.71 specificity, and an AUC of 0.80. Shapley Additive Explanations analysis revealed vocal improvements corresponding to clinical recovery, characterized by transitions from unstable, noisy voice patterns to more stable, harmonic-rich, brighter, and stronger vocal expressions. Cross-device comparison demonstrated high consistency among recordings across different smartphone brands, supporting the feasibility of mobile-based voice data collection. CONCLUSION: We developed an explainable, high-performing voice-based HF monitoring model deployable via smartphones. The approach is non-invasive, low-burden, and feasible for integration into nurse-coordinated post-discharge monitoring and remote triage.