RetiSEM: Generalising Causal Models for Fragmented Biomedical Data

In biomedical research, clinical, molecular, and imaging data are often collected in separate cohorts, making it difficult to learn causal relationships. The authors address the challenge of performing causal discovery and mediation analysis when these multimodal data are fragmented and not jointly observed at the participant level.

The authors propose RetiSEM, a framework that incorporates domain knowledge into structural equation modeling (SEM). It organizes variables into a hierarchical, biologically informed structure: genetic/ancestry-proxy variables, molecular traits, retinal microvascular features, and vascular outcomes. By applying a 'forbidden-edge' mask, the model prevents biologically implausible relationships (e.g., outcomes causing genetics) and reduces the search space for causal discovery. The framework then uses causal mediation analysis to decompose effects into total, direct, and indirect components, allowing researchers to determine if retinal features act as passive indicators or active mediators in disease pathways.

RetiSEM was evaluated against several baseline causal discovery methods (such as PC, NOTEARS, and DAGMA) across ten synthetic benchmarks that varied in dimensionality, nonlinearity, and causal depth. The framework consistently achieved lower structural Hamming distance (SHD) and higher causal accuracy than unconstrained baselines. In a real-world analysis using NHANES data combined with retinal representations, the model identified that retinal variables primarily function as downstream biomarkers rather than primary causal drivers of vascular outcomes.

This work provides a practical, interpretable tool for researchers working with limited-resource biomedical data. By moving beyond black-box predictive models, RetiSEM allows for the testing of structured causal hypotheses, helping to bridge the gap between associative deep learning predictions and the underlying biological mechanisms of systemic vascular disease.