Background Pathological changes leading to Alzheimer’s disease (AD) begin several years prior to symptom onset, underscoring the need for scalable, low-burden biomarkers to assist in early detection and longitudinal monitoring. Speech assessment is a promising candidate because it can be conducted remotely and reflects multiple domains may be affected along the AD continuum, including memory, executive function, language, and speech motor control.Methods We conducted a narrative review of PubMed (MEDLINE), Embase, Scopus, Web of Science, and the Cochrane Central Register of Controlled Trials (CENTRAL) for studies published between 1st January 2015 and 31st January 2026. Eligible studies applied machine learning (ML) approaches to speech-based assessment in subjective cognitive decline, mild cognitive impairment (MCI), prodromal AD, or early symptomatic AD. The evidence has been synthesized based on task design, feature family, endpoint type, and deployment context.Results 45 studies were included. Across studies, temporal and prosodic acoustic features, linguistic/discourse measures, and learned embeddings showed certain diagnostic potential for early AD and MCI. Multimodal fusion often only providing modest gains. Recent work has used larger cohorts, smartphone-based assessment, multilingual settings, and pathology-anchored outcomes, including amyloid-related endpoints. However, the evidence remains heterogeneous and mostly cross-sectional, with limited external validation and inconsistent handling of confounding variables.Conclusion Speech-based ML biomarkers show promise as scalable adjunctive tools for early AD and MCI assessment. Translation will require clinically anchored longitudinal validation, calibration-focused reporting, and stronger evaluation across sites, devices, and languages.
Jinming et al. (Tue,) studied this question.