Speech production as an artificial intelligence-based 'process' measure of cognition sensitive to mild cognitive impairment and Alzheimer's disease.

Abstract

Objective: Process scores in neuropsychological tests add incremental validity for detecting non-normative cognitive aging trajectories. However, process scores are laborious and time-consuming to derive. Using AI-driven natural language processing, we investigated objective speech markers related to speech production as a potential process score for measuring cognition, identifying mild cognitive impairment (MCI) and major neurocognitive disorder due to Alzheimer's disease (AD). Method: Older adults (n = 71; cognitively healthy; n = 29; MCI, n = 26; mild AD, n = 16) completed a brief battery of cognitive testing over the telephone, including a cognitive screener and four verbal memory tests. Six speech production features were extracted from the audio recordings of the verbal memory tests. Results: Pause times showed the highest convergence with cognitive screening performance and were best for distinguishing between people with or without MCI and with or without AD. This effect varied as a function of cognitive task. Verbal and semantic recall tasks showed the strongest effects. An "unstructured" autobiographical recall task showed negligible effects. Conclusions: AI-derived pause features in speech during verbal memory tests can serve as a process score of cognitive functioning that captures neurodegeneration, though cognitive tasks must be considered. The present findings reflect an important step forward for developing speech analysis for objectively quantifying cognitive dysfunctions in people with neurodegenerative disorders.