Acoustic Analysis and Prediction of Type 2 Diabetes Mellitus Using Smartphone-Recorded Voice Segments.
Level 4 - case-series / case-control
Case-control observational study evaluating acoustic features for diabetes status classification
PubMed 40206319 · doi:10.1016/j.mcpdig.2023.08.005
What was done
A cohort of 267 participants in India (192 nondiabetic: 79 women, 113 men; 75 with type 2 diabetes mellitus [T2DM]: 18 women, 57 men) was diagnosed according to American Diabetes Association guidelines. Participants recorded a standardized phrase up to 6 times daily over 2 weeks using a smartphone application, generating 18,465 voice recordings. Fourteen acoustic features were extracted to test differences between groups and develop 5-fold cross-validated classification models incorporating acoustic parameters, age, and body mass index (BMI) in age- and BMI-matched subsets.
What was found
Significant acoustic differences were detected between nondiabetic and T2DM participants. In women, the top predictive features were pitch (P < .0001), pitch standard deviation (P < .0001), and relative average perturbation jitter (P = .02). In men, key features were intensity (P < .0001) and 11-point amplitude perturbation quotient shimmer (apq11, P < .0001). In age- and BMI-matched samples, prediction models combining voice features with age and BMI achieved classification accuracies of 0.75 ± 0.22 for women and 0.70 ± 0.10 for men.
Why it matters
This study provides proof-of-concept that voice analysis via smartphones can capture subtle acoustic changes associated with type 2 diabetes, suggesting potential utility as an accessible, non-invasive prescreening modality.
Limits
The total sample size was modest (n = 267) with a notable sex imbalance among T2DM cases (only 18 women versus 57 men). Model accuracy showed wide variance, especially in women (±0.22). Validation was limited to internal cross-validation without an external test cohort, and the abstract does not report glycemic control levels (e.g., HbA1c) or medication status.
Cited by
- partial Diabetes can be detected from acoustic voice analysis due to spectral sound changes associated with dehydration.