MediEncoder: AI method unravels cause-effect in high-dimensional biomedical data

Causal mediation analysis traditionally splits a treatment's effect into indirect pathways through mediators and direct pathways. In modern biomedical studies, high-dimensional covariates and mediators act as noisy proxies for lower-dimensional latent processes. Existing methods often rely on sparsity, linear factor models, or ignore structural dependencies—limitations that break down with nonlinear measurements. MediEncoder overcomes this by jointly learning low-dimensional representations of covariates and mediators using a coupled encoder-decoder architecture. A novel cross-factor network links treatment and covariate representations to mediator representations, preserving structural dependencies. These learned features feed into a cross-fitted efficient influence function-based estimator that yields multiply robust and asymptotically normal estimates of natural direct and indirect effects.

In simulations, MediEncoder consistently outperforms competing dimension-reduction approaches in estimation accuracy. The team applied it to Alzheimer's Disease Neuroimaging Initiative (ADNI) data, demonstrating practical utility in a real-world high-dimensional biomedical setting. The paper (43 pages, 3 figures) is available on arXiv under classification stat.ME, cs.LG, math.ST, and stat.ML. MediEncoder opens the door to robust causal inference in complex, nonlinear, high-dimensional domains where traditional methods fall short.