A Gatekept Deep Semiparametric Strategy for Evaluating Lung Transplant Treatment Effect Heterogeneity

Abstract

Bilateral lung transplantation (BLT) can improve post-transplant lung function compared with single lung transplantation, but its additional functional benefit may not be uniform across recipients. Because BLT uses both donor lungs, transplant programs need to know whether its effect varies across clinically recognizable recipient profiles. Estimating such heterogeneity from observational registries is challenging. Donor and recipient characteristics influence both treatment selection and outcomes, and flexible individualized effect estimators can produce apparent subgroup variation even when no heterogeneity is present. We develop deepHTL, a gatekept semiparametric deep learning workflow that tests for heterogeneity, screens for the covariates driving it, and summarizes individualized and group average effects only when global evidence supports heterogeneity. deepHTL adjusts for confounding with cross-fitted bagged deep neural networks and separates the dominant average effect from residual heterogeneity through a revised estimator based on the Robinson transformation. In 14,196 adult recipients from the US national transplant registry, deepHTL finds strong evidence that the effect of BLT on forced expiratory volume in 1 second (FEV1) varies across recipients. Benefits are largest for recipients with low baseline FEV1, low lung allocation score, or low body mass index, whereas those with preserved baseline lung function gain little. Simulations under registry-like confounding show that deepHTL controls the type I error rate, identifies true effect modifiers, and estimates individualized effects accurately, supporting testing before interpretation in observational registries.