Multimodal AI for Alzheimer Disease Diagnosis: Systematic Review of Datasets, Models, and Modalities

Area of Science:

• Neuroscience
• Artificial Intelligence
• Medical Diagnostics

Background:

• Early detection of Alzheimer disease (AD) is crucial for effective intervention, but diagnostic accuracy varies significantly across different methods and datasets.
• Multimodal artificial intelligence (AI) models offer promising advancements, yet a fragmented evidence base hinders their widespread adoption due to diverse datasets and methodologies.

Purpose of the Study:

• To systematically review and analyze studies on multimodal AI models for AD diagnosis, prognosis, and risk prediction over a five-year period.
• To evaluate dataset characteristics, modality combinations, modeling strategies, performance metrics, and methodological limitations of these AI models.
• To discuss the real-world implications and translational pathways for multimodal AI in AD care.

Main Methods:

• Systematic literature search following PRISMA 2020 guidelines across major scientific databases (PubMed, IEEE Xplore, Scopus, ACM Digital Library, Cochrane, arXiv).
• Inclusion of studies using multimodal machine learning or deep learning for AD, mild cognitive impairment (MCI), and dementia outcomes, excluding single-modality or methodologically incomplete studies.
• Quality assessment using QUADAS-2 tool and synthesis of performance results across four major multimodal dataset families.

Main Results:

• Sixty-six studies met inclusion criteria, with multimodal models consistently outperforming single-modal baselines.
• High diagnostic accuracy (92.5%) reported for AD using the Alzheimer's Disease Neuroimaging Initiative (ADNI) dataset; strong performance (AUC 0.922) in predicting MCI conversion.
• Risk prediction in UK Biobank showed AUC of 0.84, speech-language analysis AUC of 0.813, and cross-lingual AD detection accuracy of 77%; self-collected datasets achieved ~96% accuracy but lack generalizability.

Conclusions:

• Multimodal AI models effectively integrate diverse data (biological, clinical, behavioral) to enhance AD diagnosis, prognosis, and risk prediction, outperforming single-modality approaches.
• Significant heterogeneity in datasets, outcome definitions, and validation methods limits the generalizability of current findings, highlighting prevalent risks of bias.
• Standardized benchmarks, transparent evaluation, and clinically integrated design are essential for reliable real-world deployment of multimodal AI for equitable and scalable AD diagnosis.