1308-P: A Voice-Based AI Algorithm Can Predict Type 2 Diabetes Status—Findings from the Colive Voice Study on U.S. Adult Participants

• Published in: Diabetes (New York, N.Y.) Vol. 73; no. Supplement_1; p. 1
• Main Authors: ELBEJI, ABIR, PIZZIMENTI, MÉGANE, AGUAYO, GLORIA A., FISCHER, AURELIE, AYADI, HANIN, MAUVAIS-JARVIS, FRANCK, RIVELINE, JEAN-PIERRE, DESPOTOVIC, VLADIMIR, FAGHERAZZI, GUY
• Format: Journal Article
• Language: English
• Published: New York American Diabetes Association 14.06.2024
• Subjects: Algorithms; Artificial intelligence; Diabetes; Diabetes mellitus (non-insulin dependent); Feasibility studies; Females; Hypertension; Males; Population studies; Public health; Risk assessment; United States > US
• ISSN: 0012-1797; 1939-327X
• DOI: 10.2337/db24-1308-P

Introduction: Reducing undiagnosed type 2 diabetes (T2D) cases worldwide is an urgent public health challenge. Most current screening methods are invasive, lab-based, and costly. Meanwhile, there is a growing focus on noninvasive T2D detection through advanced artificial intelligence (AI) and digital technology. This study explores the feasibility of using a voice-based AI algorithm to predict T2D status in adults, a preliminary step toward innovative screening tools. Objective: To develop and assess the performance of a voice-based AI algorithm for T2D status detection in the adult population in the US. Methods: We analyzed text reading voice recordings from 607 US participants from the Colive Voice study, adhering to the CONSORT AI standards. We trained and cross-validated algorithms with BYOL-S/CvT embeddings for each gender, evaluating them on accuracy, precision, recall, and AUC. Performance of the best models was stratified by age, BMI, and hypertension, and compared to the American Diabetes Association (ADA) score for T2D risk assessment using a Bland-Altman analysis. Results: We analyzed 323 females and 284 males; Females with T2D (age: 49.5 years, BMI: 35.8 kg/m²) vs without (40.0 years, 28.0 kg/m²). Males with T2D (47.6 years, 32.8 kg/m²) vs without (41.6 years, 26.6 kg/m²). The voice-based algorithm achieved good overall predictive capacity (AUC=75% for males, 71% for females) and correctly predicted 71% of male and 66% of female T2D cases. It is enhanced in females aged 60 years (AUC=74%) or older but also with the presence of hypertension for both genders (AUC=75%). We observed an overall agreement above 93% with the ADA risk score. Conclusion: This study demonstrates the feasibility of detecting T2D using exclusively voice features. It is the first step toward using voice analysis as a first-line T2D screening strategy. While the findings are promising, further research and validation are necessary to specifically target early-stage T2D cases.