Artificial Intelligence-Related Voice and Speech Biomarkers of Alzheimer's Disease: A Systematic Review.

BACKGROUND: Artificial intelligence (AI)-based voice and speech analysis is emerging as a promising, noninvasive biomarker for Alzheimer's disease (AD) and mild cognitive impairment (MCI). This study aimed to systematically review AI voice and speech models for detecting AD and MCI, and to evaluate their translational potential. METHODS: Following PRISMA guidelines, two investigators searched PubMed, Scopus, and the Cochrane Library for studies reporting diagnostic and prognostic outcomes of AI voice and speech models in AD and MCI. Data were extracted from studies applying machine learning (ML) or deep learning (DL) to human speech with quantitatively reported, clinically interpretable outcomes. Methodological quality and risk of bias were formally appraised using the PROBAST and CLAIM tools. RESULTS: Sixty-three studies were included. Of the 18,540 participants, there were 8,044 patients (mean age: 73.1 years), 9,736 controls (mean age: 67 years), and 760 with unspecified dementia diagnoses. There were 9,571 females and 8,341 males (628 unspecified). Most reported high diagnostic performance for distinguishing AD from healthy controls using prosody (n = 41), temporal (n = 38), spectral (n = 35), and linguistic (n = 21) measures, frequently achieving areas under the curve greater than 0.80. In contrast, MCI was less reliably distinguished from normal aging. AI-based speech analysis thus represents a promising and scalable digital biomarker, but the current evidence base remains heavily constrained by small, single-center datasets and a widespread lack of external validation. CONCLUSION: AI-driven speech analysis holds significant potential for the early detection of cognitive decline. However, its clinical translation is currently limited by methodological bias. Future validation efforts must shift toward large-scale, multicenter, longitudinal studies with standardized speech assessments.