Voice-based prediction of prediabetes using classical machine learning models.

Area of Science:

• Biomedical Engineering
• Computational Biology
• Metabolic Disease Research

Background:

• Prediabetes is a common condition increasing risks for type 2 diabetes and cardiovascular disease.
• Over 80% of individuals with prediabetes remain undiagnosed, highlighting a critical public health gap.
• Voice analysis offers a non-invasive screening method, with prior success in detecting hypertension and type 2 diabetes.

Purpose of the Study:

• To investigate the efficacy of voice-based machine learning models in identifying individuals with prediabetes.
• To evaluate the generalizability of these voice-based models across different populations.

Main Methods:

• Participants from India and Canada provided voice recordings via a mobile app; glycemic status was assessed via HbA1c.
• 167 acoustic features were extracted from voice samples, and sex-specific machine learning models were developed.
• Models were trained using L1-regularized logistic regression (LASSO) for feature selection and SMOTE for class imbalance, with evaluation via cross-validation and holdout testing.

Main Results:

• In cross-validation, the best female model achieved 0.78 balanced accuracy, and the best male model achieved 0.68.
• Holdout testing revealed that a male XGBoost model trained on an unbalanced dataset generalized better than the cross-validated model.
• Models demonstrated poor generalization on the independent Canada dataset, with several failing to accurately identify prediabetic participants.

Conclusions:

• Voice-based models show promise for prediabetes screening in controlled settings.
• Model performance significantly declines when applied across diverse geographic or demographic groups.
• Development of more robust and applicable screening tools requires diverse training data and population-specific model tuning.