Hypoglycemia poses serious risks for type 1 diabetes patients during daily life and sleep. Current detection relies on invasive glucose monitors or symptoms alone. In the study published in the Diabetes Care, the researchers sought a machine learning tool using only voice data from smartphones for noninvasive alerts.
Methods Use Real Voice Recordings
The team ran two clinical studies collecting 540 voice clips from 22 adults with type 1 diabetes (11 female, mean age 37.3 years, HbA1c 7.1%). Recordings captured euglycemia and hypoglycemia states. Participants read text aloud or did rapid syllable repetition tasks. ML models trained solely on voice features to spot hypoglycemia.
Results Deliver High Detection Accuracy
Models achieved area under the curve of 0.90 ± 0.12 for text reading and 0.87 ± 0.15 for syllable tasks. This noninvasive method matched performance of some wearable sensors. Voice changes during low blood sugar proved reliable across participants in controlled settings.
Conclusions Open Doors for Daily Use
Voice-based ML confirms potential to detect acute states like hypoglycemia without hardware. It could integrate into apps for real-time warnings. Larger trials will test free-living accuracy and integration with diabetes management systems.