ACOUSTIC ANALYSIS of routine patient-clinician conversations identified cognitive impairment with promising accuracy, according to evidence from a newly published diagnostic study.
Researchers found that machine learning models, trained on acoustic speech features from brief clinical interactions, could distinguish patients with cognitive impairment from those without a prior diagnosis.
Acoustic Analysis Identifies Cognitive Impairment Signals
To investigate whether routine clinical conversations could provide an opportunity for practical screening, researchers analysed audio recordings collected during primary care visits. Eligible participants included older adults, aged 55 years and above, who had no documented history of dementia or mild cognitive impairment.
Across the full cohort of 966 participants, 55% were female, the mean age was 67.2 years, and the prevalence of cognitive impairment was 21%. Cognitive impairment was defined by the participants’ Montreal Cognitive Assessment score being at least 1 standard deviation below their age and education adjusted norms.
Machine Learning Demonstrates Consistent Performance
Researchers extracted multiple 30 second speech segments from recorded consultations to generate acoustic measures. Machine learning classifiers were then trained to predict cognitive impairment status from these recordings.
Findings indicated that machine learning models based on Whisper-derived acoustic features delivered the strongest performance. Diagnostic accuracy reached an Area Under the Receiver Operating Characteristic Curve (AUROC) of 0.733 (95% CI: 0.714-0.752). The findings were also reproduced in an external validation cohort recruited in a separate city. Performance remained similar with an AUROC of 0.727 (95% CI: 0.714-0.740).
Implications for Routine Primary Care Screening
Acoustic analysis highlighted pitch, timing, and variability measures as important predictors of cognitive impairment. When applied as a screening tool in the external validation cohort, the algorithm achieved a positive predictive value of 30.4% (95% CI: 28.7%–32.1%), sensitivity value of 68.2% (95% CI: 61.8%–74.6%), and a specificity value of 63.6% (95% CI: 59.8%–67.4%).
The findings suggest that short segments of natural clinical dialogue contain measurable acoustic signals associated with cognitive impairment. Machine learning models trained on these acoustic signals could potentially support the feasibility of passive, speech-based screening.
Further research could validate these results and determine how best to implement acoustic analysis in clinical practice.
Reference
Colonel JT et al. Acoustic Analysis of Primary Care Patient–Clinician Conversations to Screen for Cognitive Impairment. JAMA Neurol. 2026; DOI: 10.1001/jamaneurol.2026.1868
Featured image: Jakub on Adobe Stock
- Author:
