Counterfactual prediction on irregular eectronic health records using transformers

Loading...
Thumbnail Image

URL

Journal Title

Journal ISSN

Volume Title

School of Science | Master's thesis

Department

Major/Subject

Mcode

Language

en

Pages

57

Series

Abstract

Electronic health records contain rich longitudinal information about patient trajectories that could enable personalized treatment decisions. However, estimating counterfactual treatment effects from observational EHR data is challenging due to time-varying confounding and the irregular, sparse nature of clinical measurements. Existing counterfactual prediction methods are primarily designed for short-term, regularly sampled scenarios and cannot properly handle the longer, irregular data characteristic of routine clinical practice. This thesis integrates the g-computation framework for causal inference with time-aware attention mechanisms for irregular time series. The model is validated on both synthetic tumor growth data with controllable sparsity patterns and a large-scale cancer patient cohort from Helsinki University Hospital. The proposed model demonstrates consistent improvements over state-of-the-art baselines in both prediction accuracy and uncertainty calibration across multiple evaluation settings. Predictions produce stable performance even at long projection horizons, producing well-calibrated uncertainty estimates, which is essential for clinical decision support.

Description

Supervisor

Lähdesmäki, Harri

Thesis advisor

Koskinen, Miika
Renkonen, Risto

Other note

Citation