Identify acoustic analysis of daily conversations between patients and doctors cognitive impairment Evidence from newly published diagnostic studies shows promising accuracy.
Researchers found that a machine learning model trained on acoustic speech features from short clinical interactions was able to distinguish patients with cognitive impairment from those without a prior diagnosis.
Identifying cognitive impairment signals through acoustic analysis
To investigate whether routine clinical conversations can provide real-world screening opportunities, researchers analyzed audio recordings collected during primary care visits. Eligible participants included older adults aged 55 years and older with no documented medical history. dementia or mild cognitive impairment.
Across the entire cohort of 966 participants, 55% were women, mean age was 67.2 years, and prevalence of cognitive impairment was 21%. Cognitive impairment was defined by a participant’s Montreal Cognitive Assessment score at least 1 standard deviation below age- and education-adjusted norms.
Machine learning demonstrates consistent performance
The researchers extracted multiple 30-second audio segments from the recorded consultations and created acoustic measurements. A machine learning classifier was then trained to predict cognitive impairment status from these recordings.
The findings showed that machine learning models based on whisper-derived acoustic features provided the strongest performance. Diagnostic accuracy reached an area under the receiver operating characteristic curve (AUROC) of 0.733 (95% CI: 0.714-0.752). This finding was replicated in an external validation cohort recruited in another city. Performance was similar, with an AUROC of 0.727 (95% CI: 0.714-0.740).
Impact on routine primary care screening
Acoustic analysis revealed that measures of pitch, timing, and variability were important predictors of cognitive impairment. When applied as a screening tool in an external validation cohort, the algorithm achieved a positive predictive value of 30.4% (95% CI: 28.7% to 32.1%), a sensitivity value of 68.2% (95% CI: 61.8% to 74.6%), and a specificity value of 63.6% (95% CI: 59.8% to 67.4%).
This finding suggests that short segments of natural clinical interactions contain measurable acoustic signals associated with cognitive impairment. Machine learning models trained on these acoustic signals may support the feasibility of passive voice-based screening.
Further research may validate these results and determine the best way to implement acoustic analysis in clinical practice.
reference
Acoustic analysis of primary care patient-clinician conversations to screen for cognitive impairment. JAMA New Roll. 2026; DOI: 10.1001/jamaneurol.2026.1868
Featured image: Jakub by Adobe Stock
