6th Edition of Neurology World Conference 2026

Abstract

Speech and language impairments are among the earliest observable indicators of dementia, making computational analysis of speech a promising, non-invasive marker for identifying cognitive decline. All-cause dementia affects lexical diversity, grammatical structure, and speech fluency, but many studies rely on limited linguistic features or have small sample sizes. This study investigated whether large-scale computational analysis of speech could predict the likelihood of an all-cause dementia diagnosis using linguistic features as predictors. Speech recordings from 2,454 participants drawn from the DementiaBank database (a section of TalkBank), including participants with all-cause dementia and cognitively unimpaired controls, were analyzed using custom-written code to extract a comprehensive set of speech and language measures. Indices included: speech quantity and fluency; linguistic complexity; lexical diversity; grammatical and morphosyntactic composition; and speech error patterns. Machine learning–based feature selection (i.e., LASSO regularization) was applied to identify the most informative predictors of dementia diagnosis. Frequency-weighted lexical diversity (FREQ_TTR) was strongly positively associated with dementia (OR=2.55, 95% CI: 2.26–2.88, p<0.001), while utterance-level errors (Utt_Errors) were associated with a lower likelihood of dementia (OR = 0.69, 95% CI: 0.62–0.76, p<0.001). Word-level error percentage (%_Word_Errors) showed a modest positive association (OR=1.22, 95% CI: 1.05–1.43, p=0.012). Open/closed-class word ratio (open_closed) was not significantly associated with dementia (OR=0.93, 95% CI: 0.80–1.08, p=0.366). The model demonstrated modest explanatory power (McFadden’s Pseudo R²=0.133) and acceptable multicollinearity (all VIFs<1.1). These findings demonstrate that detailed computational analysis of speech can reveal interpretable linguistic markers associated with all-cause dementia. Speech-based approaches for scalable, non-invasive assessment of cognitive decline may be useful in combination with other factors.