Back

A 60-Second Interpretable Voice Model for Early Dementia Screening

Mekulu, K.; Aqlan, F.; Yang, H.

2025-07-17 health informatics
10.1101/2025.07.16.25331667 medRxiv
Show abstract

Early detection of cognitive impairment in assisted living is hindered by time-intensive tools like MMSE and MoCA. We present a 60-second voice-based screening model that analyzes picture descriptions to estimate dementia risk. Using transcripts from the DementiaBank corpus, our model integrates traditional linguistic features (pause rate, pronoun use, syntactic complexity) with latent semantic dimensions extracted from language model embeddings. These semantic axes, interpretable constructs like "Drift & Hesitation" or "Over-detailed Narration", consistently emerged as top predictors and may represent novel linguistic biomarkers of early decline. The final ElasticNet classifier is sparse, interpretable, and outperforms known non-deep learning baselines (AUC = 0.858), exceeding MMSE. Its simplicity enables deployment in mobile apps or in-room monitors, offering scalable, low-burden screening for early dementia. This work supports a shift toward linguistically grounded, tech-enabled cognitive care in aging populations. Author summaryEarly-stage dementia often goes undetected in assisted living communities, where time constraints and staffing limitations make routine cognitive screening impractical. While standard tools like the MMSE and MoCA require trained administration and take 10-15 minutes, our research introduces a fast, interpretable alternative: a 60-second voice-based screening model using picture description tasks. We analyzed speech samples from individuals describing a common scene, extracting both traditional linguistic features (e.g., pauses, pronouns, sentence complexity) and deeper thematic patterns using modern language embeddings. Our model not only outperforms widely used tools in accuracy, but also reveals interpretable language patterns such as hesitation or over-description that may serve as early signs of cognitive decline. Lightweight and explainable, this approach is well suited for mobile apps or in-room monitors, enabling scalable, low-burden dementia screening in real-world care environments.

Published in PLOS Digital Health (predicted rank #11) · training set

Matching journals

The top 9 journals account for 50% of the predicted probability mass.

50% of probability mass above

"Similar papers" are the closest papers from that journal in the model's embedding space. They show what the match is built on, but the ranking comes mostly from a classifier over the whole training set, not from these examples alone.