The authors propose RetiSEM, a domain-constrained structural equation modelling framework designed to recover causal graphs and perform mediation analysis using fragmented biomedical data with limited multimodal resources. The method organizes variables into biologically informed blocks and applies forbidden-edge constraints to decompose pathway-level effects.
- Organises variables into biologically informed blocks and applies forbidden-edge constraints for causal graph recovery.
- Decomposes pathway-level effects into Total Effect (TE), Natural Direct Effect (NDE), and Natural Indirect Effect (NIE) components.
- Achieves lower structural error and higher causal accuracy than unconstrained baselines across ten synthetic benchmark scenarios.
- Demonstrates that retinal variables act mainly as downstream biomarker-like indicators with detectable indirect effects in real-world NHANES data.
This approach provides an interpretable framework for testing structured causal hypotheses in biomedical AI applications where clinical, molecular, and imaging variables are often incomplete or not jointly observed.