Multimodal, Explainable, and Adversarially-Robust Deep Learning for Alzheimer’s Disease Progression Detection

Project Description

Alzheimer’s disease (AD) is a progressive neurodegenerative disorder with no known cure. Accurate early prediction of its progression—from cognitively normal (CN) and mild cognitive impairment (MCI) to full AD—is essential for enabling timely clinical intervention.

Traditional diagnostic models, often reliant on a single modality or snapshot observations, fall short due to the disease’s heterogeneous and longitudinal nature. To address these challenges, InfoLab at Sungkyunkwan University (SKKU) launched a research initiative combining:

• Multimodal data fusion
• Deep and time-aware learning
• Visual explainability (XAI)
• Adversarial robustness evaluation

The goal is to develop robust, interpretable, and clinically deployable AI systems for AD progression detection.

Core Research Contributions

1. Multimodal and Cost-Effective Early Detection

We developed cost-efficient ML models using non-invasive, longitudinal data sources:

• Cognitive scores: MMSE, ADAS, CDRSB, FAQ
• Demographics
• Medication/comorbidity history (encoded using ATC ontology)

Models trained on time-series records across 2.5 years achieved strong performance in 4-class classification (CN, sMCI, pMCI, AD), with Random Forests outperforming others—even without neuroimaging data.

2. Multimodal Multitask Deep Learning with Temporal Awareness

We proposed hybrid deep learning architectures combining CNNs + BiLSTMs to model spatial-temporal dependencies in longitudinal health records.

These models jointly:

• Classify patient stage (CN, MCI, AD)
• Regress cognitive scores (e.g., MMSE) for fine-grained assessment

This multitask learning strategy improved accuracy and robustness on the ADNI dataset, surpassing single-modality models.

3. Explainable and Visual Deep Learning for Medical Trust

To improve clinical trust and transparency, we implemented a temporal visual XAI module using guided Grad-CAM over longitudinal 3D MRI scans.

Key Features:

• Voxel-level explanations highlighting brain region activity over time
• Tracks structural brain changes contributing to predictions
• Enhances clinician validation of model outcomes

4. Adversarial Robustness in Predictive Frameworks

Medical AI systems must remain robust under input distortion and adversarial conditions. We investigated:

• Effects of data perturbation, missing features, and input tampering
• Resilience improvements via multimodal fusion and multitask training

Our findings showed that fused inputs and shared learning goals help stabilize performance even under realistic attack scenarios.

Project Objectives

• Develop clinically acceptable deep learning systems for AD progression detection using multimodal, time-series data
• Build models that are interpretable, transparent, and compliant with regulatory guidelines
• Ensure adversarial resilience and consistent performance under imperfect data conditions
• Optimize for real-world deployment, particularly in resource-constrained medical settings
• Release datasets, interpretable modules, and evaluation frameworks to support reproducibility and collaboration

Research Impact

This project bridges the gap between cutting-edge AI and clinical utility in the diagnosis of neurodegenerative diseases. By integrating:

• Early fusion of multimodal inputs
• Temporal deep learning
• Adversarial threat modeling
• Visual explanation modules

The research from InfoLab (SKKU) sets a new benchmark for trustworthy, transparent, and deployable AI systems in medical diagnostics.