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

Area of Science:

• Neurology
• Artificial Intelligence
• Biomedical Informatics

Background:

• Accurate prediction of Alzheimer's disease (AD) progression in early Mild Cognitive Impairment (MCI) is vital for timely interventions.
• Current multimodal data prediction methods are limited by reliance on labeled data and failure to identify time-relevant features.

Purpose of the Study:

• To develop a novel semi-supervised multimodal learning framework, HIBMatch, for improved prediction of MCI to AD conversion.
• To address limitations in existing methods by incorporating hypergraph knowledge, information bottleneck, and consistency regularization.

Main Methods:

• Utilized hypergraphs to represent multimodal data (imaging and non-imaging).
• Proposed a Hypergraph Information Bottleneck (HIB) to focus on relevant features for future prediction.
• Implemented consistency regularization and cross-modal contrastive loss for enhanced robustness and data efficiency.

Main Results:

• HIBMatch demonstrated superior performance in predicting Alzheimer's disease prognosis on the ADNI dataset.
• The framework effectively harmonizes relevant multimodal information for future conversion prediction.
• Achieved state-of-the-art results, surpassing existing prediction methods.

Conclusions:

• HIBMatch offers a robust and data-efficient approach for predicting Alzheimer's disease progression from MCI.
• The novel hypergraph-based architecture enhances the accuracy and generalizability of prognostic models.
• This method holds significant potential for clinical application in early AD detection and management.