Study population characteristics.

The pressing need to reduce undiagnosed type 2 diabetes (T2D) globally calls for innovative screening approaches. This study investigates the potential of using a voice-based algorithm to predict T2D status in adults, as the first step towards developing a non-invasive and scalable screening method. We analyzed pre-specified text recordings from 607 US participants from the Colive Voice study registered on ClinicalTrials.gov (NCT04848623). Using hybrid BYOL-S/CvT embeddings, we constructed gender-specific algorithms to predict T2D status, evaluated through cross-validation based on accuracy, specificity, sensitivity, and Area Under the Curve (AUC). The best models were stratified by key factors such as age, BMI, and hypertension, and compared to the American Diabetes Association (ADA) score for T2D risk assessment using Bland-Altman analysis. The voice-based algorithms demonstrated good predictive capacity (AUC = 75% for males, 71% for females), correctly predicting 71% of male and 66% of female T2D cases. Performance improved in females aged 60 years or older (AUC = 74%) and individuals with hypertension (AUC = 75%), with an overall agreement above 93% with the ADA risk score. Our findings suggest that voice-based algorithms could serve as a more accessible, cost-effective, and noninvasive screening tool for T2D. While these results are promising, further validation is needed, particularly for early-stage T2D cases and more diverse populations.

全球亟需减少未确诊2型糖尿病（type 2 diabetes, T2D）的病例数，这一需求推动了创新筛查方法的研发。本研究探索了基于语音算法预测成人2型糖尿病患病状态的潜力，以此作为开发非侵入性、可规模化推广的筛查方法的第一步。我们对注册于ClinicalTrials.gov（试验编号NCT04848623）的Colive Voice研究中607名美国受试者的预设语音录音进行了分析。本研究采用混合BYOL-S/CvT嵌入特征，构建了性别特异性算法以预测2型糖尿病患病状态，并通过交叉验证，基于准确率、特异度、灵敏度以及曲线下面积（Area Under the Curve, AUC）对模型性能进行评估。我们依据年龄、体重指数（Body Mass Index, BMI）、高血压等关键因素对最优模型进行分层，并采用布兰德-奥特曼（Bland-Altman）分析法，将其与美国糖尿病协会（American Diabetes Association, ADA）的2型糖尿病风险评估评分进行对比。基于语音的算法展现出良好的预测性能：男性受试者的曲线下面积为75%，女性为71%，可正确预测71%的男性2型糖尿病病例与66%的女性2型糖尿病病例。在60岁及以上的女性受试者以及合并高血压的人群中，模型性能进一步提升，曲线下面积分别达到74%与75%，且与美国糖尿病协会的风险评分总体一致性超过93%。本研究结果表明，基于语音的算法可作为一种更具可及性、成本效益更高且非侵入性的2型糖尿病筛查工具。尽管上述结果颇具前景，但仍需开展进一步验证，尤其是针对早期2型糖尿病病例以及更多样化的研究人群。