Disputation Andrei Sirazitdinov Graph Neural Networks for Individual Treatment Effect Estimation

This dissertation advances the field of causal inference by developing and evaluating Graph Neural Network (GNN)-based methods for estimating Individual Treatment Effects (ITE), leveraging causal graph structures to improve predictive accuracy. Traditional ITE estimation approaches often fail to account for dependencies among covariates, limiting their performance, particularly in data-scarce scenarios. To address this, we propose two novel architectures, GNN-TARnet and GAT-TARnet, which integrate structural causal models with GNNs to explicitly model these dependencies. We evaluate the proposed methods on synthetic datasets with known causal structures, established benchmarks such as IHDP and JOBS, and real-world randomized controlled trial data from the PerPAIN consortium. PerPAIN is a German research initiative focused on developing personalized treatment strategies for chronic musculoskeletal pain. 
Our models consistently outperform non-structural baselines, achieving lower error in low-data settings while remaining competitive with state-of-the-art approaches when data is abundant. The practical application to the PerPAIN trial, which tests tailored psychological interventions based on patient pain profiles, highlights the utility of GNN-based ITE estimation in real-world treatment allocation and demonstrates superior performance compared to clustering-based strategies. Key contributions of this work include a peer-reviewed publication, open-source software, and a web application for patient stratification, bridging theoretical innovation with practical tools for personalized decision-making.