A 60-second interpretable voice model for early dementia screening
Kevin Mekulu1, Faisal Aqlan2, Hui Yang1
1Harold and Inge Marcus Department of Industrial and Manufacturing Engineering, Pennsylvania State University, University Park, Pennsylvania, United States of America.
PLOS Digital Health
|July 27, 2026
Summary
A new 60-second voice analysis tool screens for dementia risk by analyzing picture descriptions. This accessible technology offers scalable, low-burden cognitive care for aging populations.
Area of Science:
- Computational linguistics
- Geriatric medicine
- Artificial intelligence in healthcare
Background:
- Current dementia screening tools (MMSE, MoCA) are time-intensive.
- Deep learning models for cognitive impairment lack interpretability and require large datasets.
- Scalable and accessible early detection methods are needed for aging populations.
Purpose of the Study:
- To develop and validate a rapid, voice-based screening model for early dementia risk detection.
- To integrate traditional linguistic features with novel semantic dimensions for improved accuracy.
- To create an interpretable and computationally efficient model for clinical adoption.
Main Methods:
- Utilized transcripts from the DementiaBank corpus for model training.
- Combined traditional linguistic features (e.g., pause rate, pronoun use) with semantic axes from language model embeddings.
- Developed an interpretable ElasticNet classifier for dementia risk estimation.
Main Results:
- The voice-based model achieved an AUC of 0.858, outperforming the Mini-Mental State Examination (MMSE).
- Interpretable semantic dimensions like "Drift & Hesitation" emerged as significant predictors of cognitive decline.
- The model demonstrated superior performance compared to non-deep learning baselines.
Conclusions:
- A 60-second voice analysis model offers a scalable, low-burden approach for early dementia screening.
- Linguistically grounded semantic features show promise as novel biomarkers for cognitive impairment.
- The interpretable and efficient model design facilitates deployment in mobile health applications and monitoring systems.