HIBMatch: Hypergraph Information Bottleneck for Semi-supervised Alzheimer's Progression

TL;DR

The paper tackles the challenge of predicting Alzheimer's disease progression from MCI by leveraging multimodal data in a semi-supervised setting. It introduces HIBMatch, a Hypergraph Information Bottleneck framework that uses hypergraphs to model high-order relationships, an information bottleneck to distill minimal yet predictive information, and CHAD to enforce consistency with a discriminative classifier, plus CroC to exploit unlabeled data. The method demonstrates state-of-the-art performance on the ADNI dataset, showing strong robustness to label scarcity, perturbations, and cross-dataset generalisation to AIBL. The findings highlight the practicality of hypergraph-based, information-theoretic approaches for prognosis in heterogeneous clinical data scenarios and set a foundation for future extensions to incomplete modality settings.

Abstract

Alzheimer's disease progression prediction is critical for patients with early Mild Cognitive Impairment (MCI) to enable timely intervention and improve their quality of life. While existing progression prediction techniques demonstrate potential with multimodal data, they are highly limited by their reliance on labelled data and fail to account for a key element of future progression prediction: not all features extracted at the current moment may be relevant for predicting progression several years later. To address these limitations in the literature, we design a novel semi-supervised multimodal learning hypergraph architecture, termed HIBMatch, by harnessing hypergraph knowledge based on information bottleneck and consistency regularisation strategies. Firstly, our framework utilises hypergraphs to represent multimodal data, encompassing both imaging and non-imaging modalities. Secondly, to harmonise relevant information from the currently captured data for future MCI conversion prediction, we propose a Hypergraph Information Bottleneck (HIB) that discriminates against irrelevant information, thereby focusing exclusively on harmonising relevant information for future MCI conversion prediction. Thirdly, our method enforces consistency regularisation between the HIB and a discriminative classifier to enhance the robustness and generalisation capabilities of HIBMatch under both topological and feature perturbations. Finally, to fully exploit the unlabeled data, HIBMatch incorporates a cross-modal contrastive loss for data efficiency. Extensive experiments on the Alzheimer's Disease Neuroimaging Initiative (ADNI) dataset demonstrate that our proposed HIBMatch framework surpasses existing state-of-the-art methods in Alzheimer's disease prognosis.

Figures (7)

• Figure 1: (a) Hypergraph Information Bottleneck (HIB) optimises the representation $Z$ to capture the minimal sufficient information within the input data $D=(G,I)$ to predict the MCI conversion label $Y$. (b) Consistency on HIB and A Discriminative classifier (CHAD) updates the indiscriminative features, i.e., the red circles, which result in different predictions for HIB and a discriminative classifier, to be discriminative by forcing the features to achieve consistent predictions among two different classifiers. (c) Cross-modal Contrastive (CroC) loss is applied to unlabelled data so that different modalities of the same subject are closer than the other subjects.
• Figure 2: Illustration of the whole workflow for labelled data (top) and unlabelled data (bottom). For labelled data, different modalities, including T2*-weighted MRI, amyloid-PET, and non-imaging (age, gender, education years, cognitive and genetic scores), are input to different backbones for feature extraction. These features are used to construct a hypergraph which is then integrated with information bottleneck to obtain HIB. HIB is supervised by a HIB loss, $L_{HIB}$. HGNNP is a kind of hypergraph convolutional layer. The HIB and a discriminative classifier, trained using feature embeddings, are trained to predict the progression of AD. We further enforce a consistent prediction on their predictions by implementing a $L_{CHAD}$. For unlabelled data, the MRI and PET images of subject $v$ are augmented twice using different augmentation strategies, leading to four versions of feature representations. Then, the four versions are pulled closer and pushed far away from the features of the other subjects $u$ in a mini-batch of $B$ subjects. This is achieved by Cross-modal Contrastive (CroC) loss, $L_{CroC}$. Note that only subjects with all three modalities are retained.
• Figure 3: Pairwise comparisons between HIBMatch and existing methods. Each subplot shows the AUC distribution for the paired methods, with dashed lines connecting individual samples. The red and blue circles indicate median AUC scores.
• Figure 4: Statistical Analysis: (a) Violin plots of AUC scores for each method, with the Friedman test results ($\chi^2 = 7.04, W = 0.12, \text{CI}_{95\%} [0.06, 1.00]$).
• Figure 5: Ablation study on the modalities. All Three means using all three modalities, i.e., MRI, PET, and Non-imaging data.