Abstract
The widespread adoption of Electronic Health Records (EHR) has generated massive repositories of longitudinal patient data, offering unprecedented opportunities for data-driven healthcare and precision medicine. A central challenge in this domain is the estimation of Personalized Treatment Effects (PTE), which aims to predict the individual-level response to specific medical interventions. Unlike prediction tasks in standard supervised learning, PTE estimation is fundamentally a causal inference problem characterized by the absence of counterfactual outcomes in the training data. Furthermore, observational EHR data is plagued by selection bias, where treatment assignments are correlated with patient characteristics, leading to confounding. In this paper, we propose a novel deep learning framework for estimating PTE from high-dimensional, sparse, and irregular EHR data. Our approach leverages Causal Representation Learning to project patient covariates into a latent space where the distribution of control and treated populations is balanced, thereby mitigating selection bias. We introduce a disentangled representation architecture that separates instrumental factors, confounding factors, and adjustment factors, ensuring that the learned representations retain predictive power for the outcome while satisfying the ignorability assumption. We validate our method on both semi-synthetic datasets and real-world clinical cohorts derived from critical care databases. The experimental results demonstrate that our approach significantly outperforms state-of-the-art baselines in estimating heterogeneous treatment effects, providing a robust tool for clinical decision support.

This work is licensed under a Creative Commons Attribution 4.0 International License.
Copyright (c) 2025 Frank Harris (Author)