Non-invasive classification of stable HFpEF using a deep learning model trained on acoustic features of sustained vowels

BACKGROUND: Conventional identification of stable HFpEF still depends on resource-intensive clinical assessment, whereas non-invasive acoustic analysis of sustained vowels may provide a complementary and accessible classification signal. METHODS: The objective of this study was to develop and validate a classification model for stable heart failure with preserved ejection fraction (HFpEF) versus healthy controls using acoustic features extracted from the sustained vowel /ɑː/. In this retrospective case-control study, voice recordings were obtained from a primary cohort of 341 participants and an independent external validation cohort of 172 participants. The present study extracted 384 features using the Interspeech 2009 feature set and compared five deep learning models. The top-performing deep learning architecture, the multilayer perceptron (MLP), was further assessed through tenfold cross-validation, independent external validation, and feature analysis using SHapley Additive exPlanations and Local Interpretable Model-Agnostic Explanations. Additional analyses benchmarked the MLP against classical machine-learning models and examined uncertainty, calibration, prevalence sensitivity, and demographic confounding. RESULTS: For the binary classification of stable HFpEF versus healthy controls, the MLP classifier demonstrated the strongest performance in the five-model deep learning comparison, achieving a mean tenfold cross-validation accuracy of 0.8593 and an area under the curve (AUC) of 0.9130. On the independent external validation cohort, the primary MLP achieved an AUC of 0.836. In a broader benchmark, acoustic-feature models clearly outperformed age/sex-only models, and MLP remained competitive against strong classical tabular baselines. A regularized MLP sensitivity model achieved an external AUC of 0.8638 (95% bootstrap CI 0.8074-0.9188), whereas the best untuned SVM achieved an external AUC of 0.8628. Interpretability analyses highlighted MFCC-related spectral-envelope descriptors as influential feature families, although fold-wise stability supported family-level rather than single-feature conclusions. CONCLUSION: While the performance gap between cross-validation and external validation indicates room for further refinement, this study demonstrates that sustained-vowel IS09 acoustic features contain discriminative signal for stable HFpEF versus healthy controls under controlled conditions. The MLP was the strongest model among the tested deep learning architectures, but the broader benchmark showed that it was competitive rather than clearly superior to classical tabular models. These findings support acoustic-feature analysis as a preliminary research-stage adjunctive classification signal, while broader prospective validation in clinically heterogeneous populations is required before any clinical screening, triage, or monitoring role can be established.