Artificial intelligence for early diagnosis of mild cognitive impairment: A scoping review comparing diagnostic accuracy with physician performance

Abstract

Background: Mild cognitive impairment (MCI) represents a clinically critical and heterogeneous syndrome characterized by cognitive decline exceeding normal aging expectations, yet insufficient to impair daily functioning. As the prodromal stage of dementia, MCI offers a crucial intervention window during which therapeutic strategies can modify disease trajectory. Therefore, accurate and early diagnosis is of paramount importance. Artificial intelligence (AI) systems have demonstrated promising performance across multiple diagnostic modalities, yet their performance relative to that of physicians remains incompletely characterized in the literature. Methods: A systematic search of 6 electronic databases (PubMed/MEDLINE, EMBASE, Web of Science, Scopus, PsycINFO, and CINAHL) was conducted, resulting in 1358 records. Following de-duplication, systematic evaluation of abstracts and full text applying prespecified eligibility criteria, and exclusion of systematic reviews and meta-analyses, 71 primary research studies were included for analysis. Results: The included studies comprised 71 primary research studies published between 2008 and 2026. AI systems achieved diagnostic accuracy ranging from 62% to 100%, with neuroimaging-based and multimodal approaches often reporting the strongest performance (85–99%). Only 8 studies directly compared AI with physician diagnostic performance. In these limited and methodologically heterogeneous comparisons, AI systems generally matched or exceeded physician performance, although the available evidence base remains small. An artificial neural network achieved 90.0% sensitivity and 84.78% specificity compared to a panel of physicians collectively 46.66% sensitivity, while GPT-4 outperformed junior neurologists in 1 language-based study (81% vs 41–49%; P < .001). AI assistance improved neurologist diagnostic performance in 1 large study by approximately 26% in area under the receiver operating characteristic curve (AUROC) (P < .05). Most studies used internal validation without external datasets, limiting conclusions about real-world generalisability. Conclusion: AI approaches for early diagnosis of MCI appear promising and may support or enhance physician performance in selected settings, but widespread clinical adoption requires prospective validation, standardized physician benchmarking, and rigorous evaluation in real-world clinical environments.