Beyond Data Silos : Leveraging Disparate Data Sources for ITE Estimation

ICLR 2026 Conference Submission13049 Authors

Keywords: Heterogeneous Treatments Effect Estimation, Federated Learning, Healthcare

Abstract: While ML systems for Individual Treatment Effect (ITE) estimation have advanced healthcare decision-making, conventional methods require substantial amounts of costly training data for each intervention under consideration. In this work, we present a novel framework, based on causal transformers, for collaborative learning of heterogeneous ITE estimators across disparate data sources. Our approach can be deployed across distributed institutions (such as hospitals) via Federated Learning, enabling training on a large and diverse dataset (without sharing sensitive health data), and the same framework can be applied locally when multiple heterogeneous data sources exist within a single institution, breaking down data silos. The proposed method is flexible to handle diverse patient populations and non-identical patient measurements (covariates) across different data sources, while allowing for the estimation of treatment effects of disparate treatments being administered across these sources. Moreover, this framework can be utilized to predict the effects of novel and unseen treatments by utilizing available treatment level information. Thorough experimental evaluation on real-world clinical trial and widely-used research datasets demonstrates that our method surpasses existing baselines. Furthermore, analysis of our model's attention mechanisms reveals clinically meaningful disease and treatment-related patterns validated by domain expertise, demonstrating the interpretability and clinical relevance of our approach.

Primary Area: other topics in machine learning (i.e., none of the above)

Submission Number: 13049