HiSTOR: Structured Treatment-outcome Representation Learning for High-Dimensional Longitudinal Data

Abstract

HiSTOR learns structured representations of high-dimensional patient covariates and evolving treatment histories for longitudinal counterfactual prediction. The framework combines nonlinear dimension reduction with adjustment for time-varying confounding, supporting one-step and multi-step outcome prediction under alternative treatment sequences. Experiments on CVSim and CancerSim evaluate its performance across covariate dimensions, confounding strengths, and prediction horizons.

Qiao Liu

Assistant Professor of Biostatistics

My research interests include generative AI, high-dimensional data analysis, and computational biology.