A short, automated speech-based assessment detects mild cognitive impairment and dementia: a multicenter study with independent external validation

Abstract

Dementia affects more than 55 million people worldwide, and mild cognitive impairment (MCI) represents a critical window for early intervention. However, current neuropsychological assessments require specialist expertise and extended administration time, limiting their scalability for population-level screening. This study aimed to develop and externally validate a short, fully automated speech-based screening tool for cognitive impairment. In this multicenter cross-sectional study, a discovery cohort of 446 participants (153 cognitively normal [CN], 197 MCI, 96 dementia; mean age 76.3 ± 6.3 years) and an independent validation cohort of 158 participants were recruited from Silver Human Resources Centers and medical institutions. All participants completed the Cognitive Function Assessment using Automated Voice Guidance. Acoustic features—including Hidden Unit BERT-based modulation spectrum metrics (mid-band modulation energy [MBME] and high-band modulation energy [HBME]) and speech fluency indices—were extracted from recorded responses. Extreme gradient boosting models were trained with nested cross-validation on the discovery cohort, and eight task-configuration variants were evaluated in independent external validation. In internal cross-validation restricted to high-confidence diagnoses, the full model achieved an area under the curve (AUC) of 0.922 for distinguishing CN from those with MCI or dementia and 0.888 for distinguishing CN from those with MCI. A parsimonious model using only temporal orientation and immediate recall features (first and second trials) achieved a comparable AUC of 0.925 for CN versus MCI/dementia, demonstrating that delayed recall was not required for screening. Cognitive decline was characterized by decreased MBME and increased HBME during high-load tasks, consistent with cognitive-motor interference. In an independent external validation cohort restricted to high-confidence diagnoses, the parsimonious model achieved AUCs of 0.936 for CN versus MCI/dementia and 0.900 for CN versus MCI. A short, automated speech-based assessment integrating task scores with acoustic features could effectively differentiate among individuals with CN, MCI, and dementia. Its consistent performance in an independent external cohort supports potential deployment in primary care, community, and home-based settings.