JMIR Publications

ABSTRACT

Background:

Artificial intelligence (AI) models are increasingly used to support the diagnosis and prognosis of cognitive and neurodegenerative disorders. However, strong discrimination within development datasets does not establish transportability, reliable probability estimates, clinical benefit, equitable performance, or readiness for implementation.

Objective:

This systematic review evaluated AI models for the early detection and prognosis of cognitive and neurodegenerative disorders, with emphasis on validation, calibration, clinical utility, interpretability, equity, methodological quality, reporting completeness, and readiness for real-world clinical use.

Methods:

PubMed/MEDLINE, Embase, Scopus, Web of Science Core Collection, and IEEE Xplore were searched from inception through December, 2025. Backward and forward citation searching supplemented the electronic searches. Peer-reviewed human studies that developed or evaluated AI or machine-learning models for diagnosis, progression, or prognosis were eligible. Two reviewers independently performed screening and data extraction. Risk of bias was evaluated using PROBAST+AI-informed domains, and reporting completeness was assessed across prespecified core TRIPOD+AI domains. Owing to substantial clinical and methodological heterogeneity, findings were synthesized narratively.

Results:

Twenty studies met the inclusion criteria. Sixteen (80%) primarily addressed Alzheimer disease or mild cognitive impairment, and 18 (90%) incorporated neuroimaging. Independent external validation was reported in 7 studies (35%), whereas no study reported a formal calibration assessment. Four studies (20%) reported evidence extending beyond discrimination: 1 used decision-curve analysis, 2 included clinician comparisons, and 1 evaluated a prognostic index against longitudinal outcomes. Interpretability or feature-attribution methods were reported in 8 studies (40%), subgroup or fairness analyses in 2 (10%), and routine-clinical or independent hospital evaluation in 2 (10%). No study prospectively evaluated AI-guided care. Nineteen studies were judged to have high overall risk of bias, and 1 had unclear risk.

Conclusions:

The evidence remains concentrated at the model-development stage. Although promising performance and several examples of external validation were identified, calibration, net-benefit evaluation, equity assessment, prospective workflow testing, and lifecycle governance were rarely addressed. Future studies should use representative multicenter populations, preserve independent test cohorts, report calibration and decision-analytic performance, prespecify subgroup analyses, and evaluate AI-assisted care prospectively within intended clinical workflows. Clinical Trial: PROSPERO; CRD420261295103