A causal bidirectional selective state space model for imaging genetics in neurodegenerative diseases

Area of Science:

• Neuroscience
• Genetics
• Artificial Intelligence

Background:

• Brain imaging genetics seeks to understand brain disease mechanisms and improve diagnosis, especially for neurodegenerative disorders.
• Deep learning has improved feature extraction but faces challenges with long genetic sequences and causal inference.
• Establishing causal links between genetics, imaging, and disease is crucial for accurate diagnosis.

Purpose of the Study:

• To propose a deep causal bidirectional selective state space model (CausalMamba) for brain imaging genetics.
• To integrate multi-level feature extraction and causal inference for unified representation learning.
• To address challenges in extracting information from long genetic sequences and establishing causal relationships.

Main Methods:

• CausalMamba utilizes localized feature extraction from genetic and imaging data.
• A causal inference strategy with counterfactual reasoning and contrastive learning identifies relevant features and constructs a causal chain.
• A bidirectional selective state space model (BiMamba) integrates features for disease diagnosis.

Main Results:

• The model achieves 80.5% accuracy for Alzheimer's disease and 77.3% for Parkinson's disease using only genetic data.
• Demonstrates relative improvements of 4.2% and 2.7% over state-of-the-art methods.
• Shows computational efficiency and identifies causally relevant biomarkers across genome and brain.

Conclusions:

• CausalMamba effectively integrates genetic and imaging data for improved neurodegenerative disease diagnosis.
• The model successfully establishes causal relationships, enabling diagnosis with genetic data alone.
• This approach holds promise for identifying key biomarkers and advancing brain disease research.