Multimodal AI for Alzheimer’s Disease Diagnosis: A Systematic Review of Datasets, Models and Modalities (Preprint)

<sec> <title>BACKGROUND</title> Alzheimer’s disease (AD) is a major global health challenge, characterised by progressive cognitive decline and rising prevalence due to population ageing. Despite increased awareness, a substantial diagnostic gap remains, underscoring the urgent need for scalable and accurate early detection methods. Artificial intelligence (AI) provides a promising solution by integrating multimodal data to enhance diagnostic accuracy and clinical decision-making. However, current research remains fragmented, with studies often focusing separately on clinical phenotyping or linguistic-based cognitive impairment data. </sec> <sec> <title>OBJECTIVE</title> This review aims to systematically synthesise recent advances in multimodal AI approaches for AD diagnosis and risk prediction, bridging the gap between clinical phenotyping and linguistic-based cognitive assessment research. </sec> <sec> <title>METHODS</title> Following the PRISMA 2020 guidelines, a comprehensive search was conducted across major databases to identify studies published in the past five years that applied multimodal AI models for AD diagnosis or cognitive impairment detection. A total of 97 eligible studies were included for qualitative synthesis. Extracted data covered modality combinations, AI modelling strategies, evaluation metrics, and dataset usage. </sec> <sec> <title>RESULTS</title> Multimodal AI models consistently outperformed single-modality approaches in diagnostic accuracy and generalisability. Deep learning and ensemble learning frameworks were the most frequently adopted methods, offering strong feature-fusion capabilities. However, methodological heterogeneity, data imbalance, and limited dataset accessibility were recurring challenges. The review further summarises representative modality pairings such as imaging + clinical, imaging + genetic, and speech + cognitive features and provides an overview of publicly available multimodal datasets to facilitate reproducibility and benchmarking. </sec> <sec> <title>CONCLUSIONS</title> This systematic review provides a comprehensive overview of multimodal AI models for AD diagnosis, uniquely integrating both clinical phenotyping and linguistic-based cognitive assessment perspectives. It highlights key methodological trends, dataset limitations, and opportunities for future research toward standardised, interpretable, and generalisable AI-driven diagnostic frameworks for Alzheimer’s disease. </sec>