AI-Based Speech Analysis May Flag Cognitive Impairment

Machine-learning models trained on routine patient-physician conversations achieved similar performance in development and external validation cohorts in predicting cognitive impairment.

By
Andrea Surnit
June 19, 2026
4 min

Conexiant

Objective:

To evaluate the feasibility of using machine-learning analysis of routine conversations to identify cognitive impairment in older patients.

Approach:

Key Findings:

The highest-performing model achieved an area under the receiver operating characteristic curve (AUROC) of over 0.73 in both cohorts.
In the validation cohort, the best-performing algorithm had a positive predictive value of 30%, sensitivity of 68%, and specificity of 64%.
Cognitive impairment prevalence in the study population was 21%.
Deep neural network-derived acoustic features outperformed expert-defined acoustic measures.
Greater variability in pause duration and increased energy in unvoiced speech were linked to higher likelihood of cognitive impairment.

Interpretation:

Machine-learning models may identify at-risk patients without dedicated screening tasks.

Limitations:

Data were collected within two affiliated health systems, limiting generalizability.
Cognitive impairment was defined using MoCA rather than comprehensive neuropsychological assessment.
The models analyzed only acoustic characteristics without incorporating lexical or semantic content.