Multimodal biomarker AI techniques for early neurocognitive disorder diagnosis: A systematic review

Early diagnosis of Alzheimer's disease (AD) and related dementias remains challenging because no single biomarker sufficiently captures the complex and multifactorial nature of the underlying pathology. In recent years, multimodal artificial intelligence (AI) models capable of integrating heterogeneous data sources-such as neuroimaging, fluid biomarkers, genetics, and cognitive assessments-have emerged as a promising strategy to improve early detection and risk stratification.

We performed a PRISMA-guided systematic review (PROSPERO: CRD420251049848) of studies published from 2010 to 2025. We included 27 peer-reviewed studies applying AI/ML to ≥2 biomarker modalities for diagnostic classification or prognostic prediction (e.g., MCI-to-AD conversion), with an explicit emphasis on multimodal designs that incorporated at least one minimally invasive and/or widely deployable modality (e.g., cognitive tests, blood-based biomarkers, APOE/genetics, retinal imaging, or routine clinical features). Risk of bias was assessed using QUADAS-2.

Across the 27 included studies, multimodal AI models generally outperformed the best unimodal baselines, particularly when combining complementary biological information (e.g., imaging with molecular or clinical features). Diagnostic tasks more often achieved high discrimination (frequently AUCs in the ~0.85-0.95 range under internal validation), whereas prognostic prediction-especially MCI-to-AD conversion-remained more challenging (typically ~0.75-0.85 AUC in the best-performing models). However, evidence for generalizability was limited, as external validation was uncommon and QUADAS-2 frequently highlighted concerns in the Index Test domain related to overfitting risk and incomplete validation.

Overall, multimodal AI provides a more comprehensive representation of AD/MCI-related pathology than unimodal approaches and can improve early diagnostic classification and, to a lesser extent, prognostic prediction. However, translation to clinical practice is still constrained by limited external validation and heterogeneous reporting, which hamper generalizability and clinical trust. Future work should prioritize prospective multi-center studies, robust external validation, and transparent reporting (including interpretability analyses) to support real-world deployment.