Background: Cognitive impairment and dementia are major global health challenges, yet early

identification remains inconsistent across clinical and community settings. Conventional

screening tools often require substantial time and staffing and perform variably across linguistic,

educational, and cultural contexts. AI/ML-based tools may enable scalable, low burden cognitive

screening using speech, gait, sensor-derived signals, and electronic health record data. We aim

to systematically review AI/ML-based tools for cognitive screening and early detection in adults,

focusing on performance by cognitive stage, modality, and study setting.

Methods: Following PRISMA guidelines, we searched PubMed, Web of Science, and Embase for

peer-reviewed studies (2010-2025) evaluating AI/ML-enabled cognitive screening tools in

adults ≥ 18. Eligible studies recruited participants from community, primary care, outpatient, or

memory-clinic settings. Studies were required to report screening-relevant outcomes using an

appropriate comparator. We extracted data on study design, population, AI modality, reference

standard, and performance or implementation outcomes. Findings were synthesized

descriptively due to heterogeneity in study metrics.

Results: Of 8,475 studies identified, 29 studies met inclusion criteria. Across modalities, AI-

enabled tools demonstrated moderate to high discrimination (AUC 0.70-0.99) for cognitive

impairment, with the strongest performance observed when distinguishing cognitively

unimpaired individuals from dementia. Performance for earlier stages, including mild cognitive

impairment and subjective cognitive decline, was more variable. Speech and language-based

models and digitally administered cognitive tests with automated scoring were the most

frequently evaluated modalities and demonstrated consistent screening performance across

diverse populations and settings. Multimodal and sensor-based approaches (e.g. gait and eye 

tracking, motor function tests, EHR-based ML) leveraged non-traditional digital signals but were

evaluated in fewer studies with smaller samples. Only one study evaluated clinical outcomes in

routine care, reporting higher rates of incident ADRD diagnoses and diagnostic assessments

using a combined patient-facing and EHR-based approach, whereas the EHR-based algorithm

alone did not improve outcomes relative to usual care.

Conclusion: AI/ML-based screening tools show promising potential for detecting cognitive

impairment, including mild cognitive impairment and dementia. However, evidence supporting

earlier-stage detection and real-world clinical impact remains limited. These findings highlight

the need for standardized reporting, longitudinal validation, and multi-site external evaluation

to support translation into routine care.