Researchers have found in an acoustic analysis of doctor-patient conversations in primary care settings that short segments of speech contain vocal markers associated with cognitive impairment. Researchers developed a model using acoustic features from these interactions and achieved moderate sensitivity and specificity in identifying patients with previously undiagnosed cognitive impairment. Key predictors included measures of pitch, speech timing, and variability in speech patterns. These findings suggest that routine clinical conversations may provide a simple, noninvasive opportunity for early detection of cognitive decline, potentially enabling earlier assessment and intervention. The study was published in JAMA Neurology by Joseph T. and colleagues.
Investigators carried out a multi-center prospective validation study between August 2020 and December 2021, with data analysis being conducted from January 2025 to June 2025. In the investigation, a microphone together with a portable device was used for recording the conversation between the patient and the physician in routine unstructured primary care visits. The developmental stage of this investigation took place at primary care settings in New York, NY, among English-speaking adults above the age of 55 years and who had no medical record or evidence of having mild cognitive impairment or dementia in their past medical history.
For validating the geographical generalizability of the investigation, an external independent validation group with the same eligibility criteria was also selected at primary care centers located in Chicago, IL. Several 30-second clips from the recordings were extracted and acoustic biomarkers were obtained using sophisticated artificial intelligence foundation models including Whisper, HuBERT, wav2vec 2.0, and other linguistic approaches including eGeMAPS and prosody monitoring. The primary outcome, cognitive impairment, was measured using the gold standard Montreal Cognitive Assessment (MoCA).
Key findings:
- The study was successful in recruiting and analyzing an aggregate of 966 older adults across both of the main cities where testing occurred.
- The cohort in which the core models were developed using New York-based primary care patients who spoke English consisted of 787 patients.
- The independent external validation group in Chicago consisted of 179 extra patients.
- In the whole multi-center study group, 530 participants (55%) were female, the mean age measured at 67.2 years (SD = 8.1 years), and the rate of cognitive impairment prevalence was estimated at 21%.
- Models developed based on Whisper-based acoustic features demonstrated the best diagnostic accuracy with an AUROC of 0.733 (95% CI, 0.714–0.752) and the Fmax score of 0.502 (95% CI, 0.471–0.533).
- The generalization of the predictive models to the Chicago external site with similar diagnostic parameters yielded the validated AUROC of 0.727 (95% CI, 0.714–0.740) and the Fmax of 0.459 (95% CI, 0.441–0.477).
- Using the algorithm as a screening test on the holdout cohort resulted in a PPV of 30.4% (95% CI, 28.7%–32.1%), sensitivity of 68.2% (95% CI, 61.8%–74.6%), and specificity of 63.6% (95% CI, 59.8%–67.4%).
In conclusion, the findings of this diagnostic study have shown that machine learning algorithms trained with acoustic properties from short clinical dialogue segments were capable of diagnosing CI with high accuracy. Such findings indicate the possibility of passive screening using voice data obtained from normal doctor-patient communication.
Reference:
Colonel JT, Becker J, Chan L, et al. Acoustic Analysis of Primary Care Patient–Clinician Conversations to Screen for Cognitive Impairment. JAMA Neurol. Published online June 15, 2026. doi:10.1001/jamaneurol.2026.1868
Disclaimer: This website is primarily for healthcare professionals. The content here does not replace medical advice and should not be used as medical, diagnostic, endorsement, treatment, or prescription advice. Medical science evolves rapidly, and we strive to keep our information current. If you find any discrepancies, please contact us at corrections@medicaldialogues.in. Read our Correction Policy here. Nothing here should be used as a substitute for medical advice, diagnosis, or treatment. We do not endorse any healthcare advice that contradicts a physician's guidance. Use of this site is subject to our Terms of Use, Privacy Policy, and Advertisement Policy. For more details, read our Full Disclaimer here.
NOTE: Join us in combating medical misinformation. If you encounter a questionable health, medical, or medical education claim, email us at factcheck@medicaldialogues.in for evaluation.