A review of Alzheimer's disease diagnosis and prognosis models based on multimodal deep learning.

Although some drugs have been approved for clinical treatment, early diagnosis and intervention remain the most effective strategies for managing Alzheimer's disease (AD) at present. With advances in deep learning and multimodal fusion, an increasing number of complex frameworks have been proposed. This paper systematically reviews multimodal deep learning-based models for AD diagnosis between 2020 and 2026. Beyond the technical survey, we explore how to deal with the heterogeneous modality integration and missing modality processing. From these experimental results, many models show impressive performance on public datasets. However, we have noticed a troubling problem that these "lab-perfect" results often falter when they face the chaos of the real-world. Because of the persistent black-box problem and the hidden traps of data leakage, the path to clinical use is still uphill. This work suggests that it is time to move beyond chasing tiny gains in accuracy and focus on building models that doctors can truly trust, understand, and use in real clinical settings.