The authors propose a latent-variable model for estimating treatment effects from single-arm trials with external controls, focusing on scenarios where outcome information is limited. Their approach involves learning group-specific and shared latent representations to improve treatment effect estimation.
KG-TREAT introduces a novel pre-training and fine-tuning framework that synergizes patient data with biomedical knowledge graphs to enhance Treatment Effect Estimation. The approach constructs dual-focus KGs and integrates a deep bi-level attention synergy method for in-depth information fusion.
Regression's weighting problem, arising from unmodeled heterogeneity in treatment effects, can be effectively addressed by adopting the separate linearity assumption and employing established methods like regression imputation, interacted regression, or balancing weights.