[Paper Review] DeepAD: A Robust Deep Learning Model of Alzheimer's Disease Progression for Real-World Clinical Applications

Somaye Hashemifar, Claudia Iriondo+3 more|arXiv (Cornell University)|Mar 17, 2022

Machine Learning in Healthcare4 citations

TL;DR

DeepAD is a multimodal, multi-task deep learning model that predicts Alzheimer’s disease progression using longitudinal clinical, demographic, and 3D MRI data. By integrating 3D CNN-extracted imaging features with tabular data and applying adversarial loss and Sharpness-Aware Minimization, it achieves state-of-the-art generalization across diverse cohorts, reducing prediction error by up to 30% on out-of-study datasets.

ABSTRACT

The ability to predict the future trajectory of a patient is a key step toward the development of therapeutics for complex diseases such as Alzheimer's disease (AD). However, most machine learning approaches developed for prediction of disease progression are either single-task or single-modality models, which can not be directly adopted to our setting involving multi-task learning with high dimensional images. Moreover, most of those approaches are trained on a single dataset (i.e. cohort), which can not be generalized to other cohorts. We propose a novel multimodal multi-task deep learning model to predict AD progression by analyzing longitudinal clinical and neuroimaging data from multiple cohorts. Our proposed model integrates high dimensional MRI features from a 3D convolutional neural network with other data modalities, including clinical and demographic information, to predict the future trajectory of patients. Our model employs an adversarial loss to alleviate the study-specific imaging bias, in particular the inter-study domain shifts. In addition, a Sharpness-Aware Minimization (SAM) optimization technique is applied to further improve model generalization. The proposed model is trained and tested on various datasets in order to evaluate and validate the results. Our results showed that 1) our model yields significant improvement over the baseline models, and 2) models using extracted neuroimaging features from 3D convolutional neural network outperform the same models when applied to MRI-derived volumetric features.

Motivation & Objective

• To develop a robust, generalizable deep learning model for predicting Alzheimer’s disease progression across diverse clinical cohorts.
• To address the challenge of inter-study domain shift in neuroimaging data caused by scanner and protocol variations.
• To improve prediction accuracy for continuous cognitive scores (CDRSB, MMSE, ADAS-COG12) by fusing multimodal data including MRI, clinical, and demographic features.
• To enhance model generalization using adversarial loss and Sharpness-Aware Minimization (SAM) for improved robustness to distribution shifts.
• To enable personalized, fine-grained prognosis of disease trajectories for early-stage AD patients in real-world clinical settings.

Proposed method

• A 3D convolutional neural network (3D-CNN) is trained end-to-end on full-resolution T1-weighted MRI scans to extract spatiotemporal features.
• Clinical, demographic, and genomic data are processed as tabular inputs and fused with deep imaging features in a multimodal, multi-task learning framework.
• An adversarial loss is applied during training to align feature distributions across different study sites, reducing domain shift effects.
• Sharpness-Aware Minimization (SAM) is used as an optimization technique to improve generalization by seeking flatter minima in the loss landscape.
• The model is trained and evaluated on multiple datasets from the ADNI cohort, including in-study and out-of-study test sets to assess generalization.
• End-to-end training is performed to jointly optimize predictions for three cognitive endpoints: CDRSB, MMSE, and ADAS-COG12.

Experimental results

Research questions

• RQ1 Can a multimodal deep learning model that fuses clinical, demographic, and 3D MRI data improve prediction accuracy for Alzheimer’s disease progression?
• RQ2 How effective is adversarial loss in mitigating inter-study domain shift in multi-site neuroimaging data?
• RQ3 Does incorporating Sharpness-Aware Minimization (SAM) enhance model generalization across diverse clinical cohorts?
• RQ4 Does using raw 3D MRI features from a 3D-CNN outperform traditional volumetric features derived from segmentation pipelines?
• RQ5 To what extent does the model generalize to unseen study sites not included in training?

Key findings

• DeepAD achieved a 30% relative improvement in prediction accuracy on out-of-study test sets compared to baseline models, with the lowest test loss of 0.4 when using SAM optimization.
• The model using 3D-CNN extracted features outperformed models using MRI-derived volumetric features, reducing test error by up to 30% on MMSE prediction.
• Adversarial loss improved generalization, yielding a 0.22% increase in R² for MMSE prediction on study E and 0.09% on study F, which were not used in training or validation.
• Training with SAM+SGD led to more stable convergence and a 0.11 improvement in R² (from 0.16 to 0.27) compared to standard SGD.
• The model demonstrated strong robustness to domain shift, with out-of-study performance consistently outperforming in-study baselines on all three cognitive scores.
• DeepAD achieved the best performance across all metrics when combining MRI and clinical modalities, confirming the value of multimodal fusion.