Medical researchers in Canada trained a machine learning-based AI to identify 14 differences in voice between people with type 2 diabetes and non-diabetics.
These characteristics include pitch, intensity, and other small sound changes that are indistinguishable from the human ear.
Type 2 diabetes is then diagnosed by listening to the patient’s voice for 6 to 10 seconds.
Main principles:
1. Vocal Signature Recognition: AI models are trained to recognize vocal characteristics associated with type 2 diabetes. These characteristics include pitch, intensity, and other small sound changes that are indistinguishable from the human ear.
2. Machine learning training: Canadian medical researchers used 267 voice recordings from Indian residents to train AI. Of these, about 72% of participants were diagnosed with non-diabetes, and the rest were diagnosed with type 2 diabetes. All participants recorded a phrase six times a day for two weeks, resulting in a total of 18,000 recordings.
3. Analysis of voice differences: Scientists have identified 14 differences in voice between people with type 2 diabetes and non-diabetics. Among them, four differences help AI diagnose type 2 diabetes more accurately.
4. Combine other health data: In addition to sound data, AI also combines basic health data collected by researchers, such as age, gender, height, and weight, to improve the accuracy of diagnosis.
Predict the effect:
1. Diagnostic accuracy: AI can accurately diagnose type 2 diabetes in 89% of women and 86% of men. The study found that pitch and standard deviation of pitch were useful features for diagnosing diabetes in all participants.
2. Gender differences: For women, the predictive characteristics are average pitch, pitch SD, and RAP jitter. For men, average strength and APQ11 SHIMMER are used. Simply put, these characteristic changes suggest that women with type 2 diabetes report slightly lower pitches and smaller changes, while men with type 2 diabetes report slightly weaker voices with greater variations.
3. Potential applications: Researchers believe that sound analysis shows potential as a pre-screening or monitoring tool for type 2 diabetes, especially when combined with other risk factors associated with the condition.
Detailed report: https://diabetes.co.uk/news/2023/oct/say-what-ai-can-diagnose-type-2-diabetes-in-10-seconds-from-your-voice.html
The study was published in the journal Mayo Clinic Proceedings: Digital Health. https://mcpdigitalhealth.org/article/S2949-7612(23)00073-1/fulltext