Deep Learning-Based Interpretable Multimodal Data Fusion for Early Diagnosis of Alzheimer’s Disease

As the global population ages, Alzheimer's disease (AD) poses an increasing threat to public health. Early diagnosis is critical to enable timely intervention, monitoring, and trial recruitment before severe neurodegeneration occurs. Machine learning (ML) can assist clinicians with diagnostic decision-making and aid researchers in identifying biologically meaningful biomarkers. However, current diagnostic tools face major challenges: high-dimensional, small-sample gene expression data; limited interpretability of deep learning models; and the loss of complementary information during heterogeneous multimodal data fusion.

This thesis addresses these challenges through three main contributions. First, a stacked sparse autoencoder (SSAE) is developed for gene expression dimensionality reduction. By incorporating sparsity constraints and stacking hidden layers, the SSAE effectively models non-linear relationships, mitigates overfitting, and outperforms traditional dimensionality reduction algorithms in early AD classification.

Second, an interpretable feature selection framework is proposed, combining a shallow sparse autoencoder (AE) with XGBoost-based feature importance ranking. This approach captures complex non-linear data structures while retaining direct mapping between hidden nodes and input probes. By providing transparent feature selection, it addresses the black-box limitation of conventional deep learning and identifies gene expression probes with validated biological relevance through enrichment analysis.

Third, a novel heterogeneous multimodal data fusion method is introduced to integrate neuroimaging and gene expression data. Unlike existing fusion techniques that focus primarily on cross-modal consistency, this framework preserves modality-unique complementary information to construct a richer fused representation, outperforming both unimodal baselines and consistency-only multimodal methods.

In summary, this thesis presents an interpretable deep learning framework for multimodal data fusion, improving both diagnostic accuracy and biological interpretability in early AD diagnosis. Future work will focus on validating the framework across larger multi-cohort datasets, incorporating additional omics modalities, and enhancing computational efficiency and model robustness for clinical decision support.