Dementia risks identified by vocal features via telephone conversations: A novel machine learning prediction model

Area of Science:

• Neurology and Artificial Intelligence
• Biomedical Engineering
• Gerontology

Background:

• Current Alzheimer's disease (AD) diagnosis methods are costly and lack accessibility for early detection.
• Preclinical AD risk identification requires low-cost, reliable tools.
• Cognitive decline in AD may manifest in vocal characteristics during everyday speech.

Purpose of the Study:

• To develop and evaluate a novel machine learning (ML) model for identifying preclinical Alzheimer's disease (AD) risk using vocal features from daily conversations.
• To compare the predictive performance of the ML vocal analysis model against the Japanese version of the Telephone Interview for Cognitive Status (TICS-J).

Main Methods:

• Collected 1,616 audio recordings from 99 healthy controls (HC) and 24 AD patients.
• Extracted vocal features and developed ML models (XGBoost, Random Forest, Logistic Regression) for AD risk prediction.
• Evaluated model performance using ROC curves, AUCs, sensitivity, and specificity, comparing audio-based and participant-based predictions with TICS-J.

Main Results:

• Participant-based prediction models achieved high Area Under the Curve (AUC) values: XGBoost (1.000), Random Forest (1.000), and Logistic Regression (0.972).
• The ML models showed comparable or superior predictive performance to the TICS-J (AUC: 0.917).
• The novel vocal feature analysis demonstrated significant potential for AD risk assessment.

Conclusions:

• Vocal features extracted from daily conversations can be effectively utilized by machine learning models for Alzheimer's disease (AD) risk assessment.
• This approach offers a potentially low-cost, accessible, and reliable tool for preclinical AD detection.
• Further validation is warranted to integrate this technology into clinical practice for early AD risk stratification.