Sound as a bell: a deep learning approach for health status classification through speech acoustic biomarkers.

Background: Human health is a complex, dynamic concept encompassing a spectrum of states influenced by genetic, environmental, physiological, and psychological factors. Traditional Chinese Medicine categorizes health into nine body constitutional types, each reflecting unique balances or imbalances in vital energies, influencing physical, mental, and emotional states. Advances in machine learning models offer promising avenues for diagnosing conditions like Alzheimer's, dementia, and respiratory diseases by analyzing speech patterns, enabling complementary non-invasive disease diagnosis. The study aims to use speech audio to identify subhealth populations characterized by unbalanced constitution types.

Methods: Participants, aged 18-45, were selected from the Acoustic Study of Health. Audio recordings were collected using ATR2500X-USB microphones and Praat software. Exclusion criteria included recent illness, dental issues, and specific medical histories. The audio data were preprocessed to Mel-frequency cepstral coefficients (MFCCs) for model training. Three deep learning models-1-Dimensional Convolution Network (Conv1D), 2-Dimensional Convolution Network (Conv2D), and Long Short-Term Memory (LSTM)-were implemented using Python to classify health status. Saliency maps were generated to provide model explainability.

Results: The study used 1,378 recordings from balanced (healthy) and 1,413 from unbalanced (subhealth) types. The Conv1D model achieved a training accuracy of 91.91% and validation accuracy of 84.19%. The Conv2D model had 96.19% training accuracy and 84.93% validation accuracy. The LSTM model showed 92.79% training accuracy and 87.13% validation accuracy, with early signs of overfitting. AUC scores were 0.92 and 0.94 (Conv1D), 0.99 (Conv2D), and 0.97 (LSTM). All models demonstrated robust performance, with Conv2D excelling in discrimination accuracy.

Conclusions: The deep learning classification of human speech audio for health status using body constitution types showed promising results with Conv1D, Conv2D, and LSTM models. Analysis of ROC curves, training accuracy, and validation accuracy showed all models robustly distinguished between balanced and unbalanced constitution types. Conv2D excelled with good accuracy, while Conv1D and LSTM also performed well, affirming their reliability. The study integrates constitution theory and deep learning technologies to classify subhealth populations using noninvasive approach, thereby promoting personalized medicine and early intervention strategies.