Causal-Augmented Source-Free Domain Adaptation With Scale-Free Transformer for Schizophrenia Classification.

Area of Science:

• Neuroimaging
• Artificial Intelligence
• Psychiatric Disorders

Background:

• Transformer models are used for schizophrenia (SZ) diagnosis from brain functional networks (BFNs) using fMRI data.
• Existing methods overlook scale-free properties, hindering the capture of BFN topology.
• Multi-site data heterogeneity necessitates source-free domain adaptation (DA) due to data sharing restrictions.

Purpose of the Study:

• To propose a source-free DA framework integrating a scale-free transformer encoder for improved SZ classification.
• To enhance robustness against multi-site data variability by leveraging intrinsic inter-regional dependencies.
• To identify discriminative temporal regions offering insights into SZ neural mechanisms.

Main Methods:

• Pre-training a transformer encoder on labeled source domains with a scale-free prior to bias attention towards hub nodes.
• Utilizing causal graph construction from inter-regional interactions and perturbing causal structures via random permutation and counterfactual interventions.
• Employing entropy minimization for joint optimization of the encoder and predictor to learn domain-invariant representations.

Main Results:

• The proposed method achieved superior performance compared to 23 other methods across two SZ datasets, with accuracies of 87.18%±0.91% and 88.39%±0.13%.
• Ablation studies confirmed the significant contributions of causal, permutation, counterfactual, and entropy minimization constraints.
• Identified discriminative temporal regions provided novel insights into the dysfunctional neural mechanisms in SZ.

Conclusions:

• The developed source-free DA framework effectively addresses challenges in multi-site BFN analysis for SZ diagnosis.
• Integrating scale-free priors and causal inference enhances model robustness and diagnostic accuracy.
• The findings offer potential for improved understanding and diagnosis of schizophrenia through advanced neuroimaging analysis.