Abstract With the growing prevalence of cognitive decline in ageing populations, accessible and scalable screening tools are essential for early intervention. This study investigated the potential of automated speech analysis as a proxy for cognitive assessment in 1003 older adults. Employing machine learning regression models, we demonstrated that linguistic and acoustic features extracted from spontaneous speech quadrupled performance compared to models using demographic information alone, when predicting cognitive domain scores. We then trained a binary classifier to identify individuals performing below normative thresholds (ROC-AUC up to 0.81), illustrating possible applications such as large-scale screening for cognitive impairment and improved participant selection for clinical trials. Finally, we evaluated our approach on an independent clinical dataset of Alzheimer’s disease (AD) patients and controls, demonstrating its generalizability. These findings highlight the clinical feasibility of speech analysis as a low-cost, non-intrusive digital biomarker for cognitive monitoring and screening.
Heitz et al. (2026) studied this question.