MON-598 Comparative Performance of Voice Biomarkers in Detecting Prediabetes and Diabetes Among Younger vs. Older Adults

Abstract Disclosure: B. Ongphiphadhanakul: None. W. Aekplakorn: None. B. Ongphiphadhanakul: None. N. Kasemkosin: None. Background: The global burden of prediabetes and type 2 diabetes (T2D) affects all age groups, screening measures can be invasive, costly, and underutilized. Voice-based screening has emerged as a novel and non-invasive supplementary method. However, age-related changes in vocal characteristics may interfere with the detection of metabolic dysregulation from voice in older adults. Objective: To determine the feasibility and comparative performance of using vocal biomarkers extracted from a brief “/ah/” recording, combined with machine learning, to detect prediabetes or diabetes in younger and older adults. Methods: Subject consisted of 1,240 volunteers from the Thailand National Examination Survey. Each participant provided a 5-second sustained “/ah/” phonation recorded via a standard cell phone. Audio data were processed using the openSMILE Python package to extract 88 acoustic parameters based on the Extended Geneva Minimalistic Acoustic Parameter Set (eGeMAPS). The PyCaret machine learning library was used to train and validate multiple classification models (normal vs. prediabetes/diabetes, defined as HbA1c > 5.7% or fasting plasma glucose > 100 mg/dL) under five-fold cross-validation. Data were expressed as mean ± SD. Results: In adults aged 20-30 (n = 137), a Naive Bayes model achieved the best performance, yielding 77.9% ± 5.2% accuracy, an AUC of 0.85 ± 0.06, recall of 78.3% ± 19.4%, and precision of 44.2% ± 9.6%. In contrast, ridge regression performed optimally in participants aged 30 and older (n = 1103), but only reached 54% ± 3% accuracy, an AUC of 0.52 ± 0.05, and 0.58 ± 0.02 for both sensitivity and recall—suggesting that age-related vocal changes may diminish the predictive power of voice biomarkers in older adults. Conclusion: These findings underscore the feasibility of using voice analysis as a non-invasive, easily scalable method for early detection of prediabetes and diabetes, particularly in younger populations. While preliminary results are promising, additional research is necessary to refine the approach for older adults, where age-related variations in vocal characteristics may mask the effect of glycemic status on voice. Presentation: Monday, July 14, 2025