A Transformer-Based Framework for Counterfactual Estimation of Antihypertensive Treatment Effect on COVID-19 Infection Risk - A Proof-of-Concept Study

Abstract

Background: Transformer-based neural networks excel in modelling high-dimensional, time-series data with complex dependencies. This proof-of-concept study applies a transformer-X-learner framework to estimate treatment effects using real-world data, using antihypertensive drug exposure and COVID-19 risk as an exemplar. Methods: We conducted a case-control study of 303,220 NHS Greater Glasgow and Clyde patients aged ≥ 40 years during the first two COVID-19 pandemic waves. Using a transformer-X-learner framework that incorporated temporal patterns in medication usage and comorbidities, we controlled for confounding effects and estimated individual and average treatment effects ACEIs, beta-blockers (BBs), calcium channel blockers (CCBs), thiazides (THZs), and statins on 180-day SARS-CoV-2 infection risk. Results: The transformer-X-learner framework outperformed traditional approaches, achieving an F1 score of 0.82 and area under the precision-recall curve (AUPRC) of 0.78. ACEIs showed a negligible overall impact on COVID-19 risk (ATE: 0.97%±5.5), while BBs (-8.3%±7.3%) and CCBs (-9.7%±8.1%) were protective. Statins (3.5%±6.1%) and THZs (4.3%±10.8%) showed slight increases in risk. Treatment effects were consistent across age, gender, and socioeconomic categories. Conclusions ACEIs do not substantially increase the risk of COVID-19 infection while the protective effects of BBs and CCBs warrant further investigation. This study highlights the potential of transformer-based causal inference models as a powerful tool for evaluating treatment safety and efficacy in complex healthcare scenarios.