An Integrated Multimodal Deep Learning Approach for Early Dementia Prediction Using ADNI Data
Contributors
Pradeep Yadav
Shashikant Gupta
Keywords
Proceeding
Track
General Track
License
Copyright (c) 2026 Sustainable Global Societies Initiative

This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International License.
Abstract
Early identification of MCI patients progressing to Alzheimer's disease (AD) is critical for early intervention, but most deep learning methods operate on a single modality or do not take into account the issue of missing modalities and disease dynamics. In this work, we propose an end-to-end multimodal framework for the early diagnosis of dementia using ADNI data. Our method consists of encoding of structural MRI, FDG-PET, and CSF, genetic, and demographic features using modality-specific neural networks, fusion of the encoded modalities using cross-modal attention and modality dropout to allow for the use of samples with partially observed modalities, and modeling disease progression by a recurrent component. We describe a leakage-proof evaluation protocol on the subject level for predicting stable versus progressive MCI and diagnosing Alzheimer's disease on the three-class problem, including external validation across ADNI phases, calibration analysis, ablation studies, and model interpretability. This paper describes the design and evaluation protocol of our method and does not contain any empirical results.