基于语音的机器学习无创检测低血糖
该研究在1型糖尿病患者中采集标准化血糖正常与低血糖状态下的语音数据,训练并评估仅基于语音特征的机器学习方法检测低血糖。结果显示该方法具有较高准确性,朗读文本的ROC曲线下面积为0.90±0.12,快速重复音节的ROC曲线下面积为0.87±0.15。
要点
- 仅基于语音数据的机器学习方法可无创检测低血糖。
- 朗读文本任务的ROC曲线下面积为0.90±0.12。
- 快速重复音节任务的ROC曲线下面积为0.87±0.15。
核心结论
- 仅基于语音数据的机器学习方法可实现无创低血糖检测,支持通过语音推断急性健康状态的潜力。
研究设计与样本
- 两项序贯临床研究,纳入1型糖尿病患者,在标准化血糖正常与低血糖状态下用智能手机采集540条语音记录。
主要发现
- 机器学习方法检测低血糖的ROC曲线下面积:朗读文本为0.90±0.12,快速重复音节为0.87±0.15。
临床怎么用
- 语音数据可作为无创低血糖检测的潜在生物标志物,但需进一步验证后方可用于临床。
以上解读依据论文摘要(全文未开放获取或许可不允许翻译转载)。
原文
OBJECTIVE: Hypoglycemia is a hazardous diabetes-related emergency. We aimed to develop a machine learning (ML) approach for noninvasive hypoglycemia detection using voice data. RESEARCH DESIGN AND METHODS: We collected voice data (540 recordings) with a smartphone in standardized euglycemia and hypoglycemia in two sequential clinical studies in people with type 1 diabetes. Using these data, we trained and evaluated an ML approach to detect hypoglycemia solely based on voice features. RESULTS: Twenty-two individuals were included (11 female, age 37.3 ± 12.4 years, HbA1c 7.1 ± 0.5%). The ML approach detected hypoglycemia noninvasively with high accuracy (area under the receiver operating characteristic curve 0.90 ± 0.12 for reading a text aloud and 0.87 ± 0.15 for rapid repetition of syllables [diadochokinetic task]). CONCLUSIONS: An ML approach exclusively based on voice data allows for noninvasive hypoglycemia detection, corroborating the potential of ML-based approaches to infer acute health states through voice.
引用
- Lehmann V, Hilpert M, Mostaani Z, et al. Listening to Hypoglycemia: Voice as a Biomarker for Detection of a Medical Emergency Using Machine Learning. Diabetes Care. 2026. PMID: 41251597. doi:10.2337/dc25-1680
供执业医师工作参考,不替代临床判断;数值与分级以指南原文为准。