A diagnostic study finds machine learning can screen for cognitive impairment by analyzing the acoustic features of routine patient-clinician conversations.
Researchers analyzed audio recordings from primary care visits of adults over 55 with no prior dementia diagnosis. They used machine learning models trained on speech features like pitch, timing, and variability to identify cognitive decline.
The models achieved an AUROC of 0.733 in initial testing and 0.727 in an external validation cohort, demonstrating consistent performance. This suggests short segments of natural dialogue contain measurable acoustic signals of impairment.
The approach could enable passive, speech-based screening during regular check-ups, though further research is needed to validate its clinical implementation.