[Paper Review] Multi-View Graph Convolutional Network and Its Applications on Neuroimage Analysis for Parkinson's Disease

Xi Zhang, Lifang He+4 more|arXiv (Cornell University)|May 22, 2018

Parkinson's Disease Mechanisms and Treatments26 references72 citations

TL;DR

The paper introduces HOSLR, a high-order sparse logistic regression model that directly handles tensor data for clinical risk prediction, solved via block coordinate descent and proximal gradient methods, with validation on fMRI and EHR datasets showing competitive performance and interpretable feature patterns.

ABSTRACT

Parkinson's Disease (PD) is one of the most prevalent neurodegenerative diseases that affects tens of millions of Americans. PD is highly progressive and heterogeneous. Quite a few studies have been conducted in recent years on predictive or disease progression modeling of PD using clinical and biomarkers data. Neuroimaging, as another important information source for neurodegenerative disease, has also arisen considerable interests from the PD community. In this paper, we propose a deep learning method based on Graph Convolutional Networks (GCN) for fusing multiple modalities of brain images in relationship prediction which is useful for distinguishing PD cases from controls. On Parkinson's Progression Markers Initiative (PPMI) cohort, our approach achieved $0.9537\\pm 0.0587$ AUC, compared with $0.6443\\pm 0.0223$ AUC achieved by traditional approaches such as PCA.

Motivation & Objective

• Motivate risk prediction when data are multi-dimensional tensors rather than vectors (e.g., medical images, EHR tensors).
• Develop a high-order sparse logistic regression framework that preserves multi-mode structure and promotes sparsity for interpretability.
• Provide an efficient optimization algorithm with convergence guarantees to train the model on tensor data.

Proposed method

• Define tensor inputs for each observation and a multilinear logistic model using mode-k weight vectors.
• Combine L1 and L2 regularization (elastic net) across all mode-specific weight vectors to induce sparsity and stability.
• Use a Block Coordinate Descent (BCD) approach to iteratively optimize each mode’s weight vector and bias.
• Solve subproblems with a Block Proximal Gradient method including a shrinkage operator for sparsity.
• Provide convergence justification via the Kurdyka–Lojasiewicz framework (referenced in the paper).
• Demonstrate the approach on two real-world medical tasks: AD risk from fMRI tensors and CHF risk from EHR matrices.

Experimental results

Research questions

• RQ1 Can HOSLR effectively predict clinical risk from tensor-valued medical data without flattening to vectors?
• RQ2 Do elastic-net regularizations on multilinear weights improve prediction accuracy and interpretability?
• RQ3 How does HOSLR compare to vector-based methods (LR, SLR, SVM) and standard multilinear/logistic approaches on real data?
• RQ4 What interpretable risk-factor patterns emerge from tensor-based coefficients in AD and CHF datasets?

Key findings

• HOSLR outperformed NN, SVM, LR, SLR, and MLR across three AD-related binary classification tasks based on fMRI data (AUC improvements shown in Fig. 2).
• In CHF prediction from EHR-derived matrices, HOSLR achieved higher average AUC compared to baseline methods (Fig. 3).
• The model yields two mode-specific weight vectors (medication and diagnosis) enabling interpretation of important feature pairs, with top diagnoses and medications highlighted in Table 1.
• The bilinear form f(X) = w_med^T X w_diag supports interpretable feature interactions between medications and diagnoses.
• Elastic-net regularization provides numerical stability and encourages joint sparsity across modes.