Dementia Prediction From Speech Signal Using Optimized Prosodic Features

Early detection of dementia, a general term for cognitive decline, is vital for enhancing patient care and management. Conventional clinical assessments and cognitive tests require expert clinicians and are time-consuming. Therefore, developing rapid and reliable in-house detection methods is highly desirable. Prior studies have identified speech-based early markers of dementia, motivating the application of machine learning for automated detection. This study investigates the effectiveness of optimized prosodic features, with a focus on pause-related metrics, for dementia classification. Experimental evaluation of 15 selected features shows that pause and formant features are the most discriminative, achieving improved classification performance over the original 45-feature set. These findings suggest that concise, targeted feature sets can enhance automated speech-based dementia detection.