Generalizability of heterogeneous treatment effects based on causal forests applied to two randomized clinical trials of intensive glycemic control.

Purpose Machine learning is an attractive tool for identifying heterogeneous treatment effects (HTE) of interventions but generalizability of machine learning derived HTE remains unclear. We examined generalizability of HTE detected using causal forests in two similarly designed randomized trials in type II diabetes patients. Methods We evaluated published HTE of intensive versus standard glycemic control on all-cause mortality from the Action to Control Cardiovascular Risk in Diabetes study (ACCORD) in a second trial, the Veterans Affairs Diabetes Trial (VADT). We then applied causal forests to VADT, ACCORD, and pooled data from both studies and compared variable importance and subgroup effects across samples. Results HTE in ACCORD did not replicate in similar subgroups in VADT, but variable importance was correlated between VADT and ACCORD (Kendall's tau-b 0.75). Applying causal forests to pooled individual-level data yielded seven subgroups with similar HTE across both studies, ranging from risk difference of all-cause mortality of -3.9% (95% CI -7.0, -0.8) to 4.7% (95% CI 1.8, 7.5). Conclusions Machine learning detection of HTE subgroups from randomized trials may not generalize across study samples even when variable importance is correlated. Pooling individual-level data may overcome differences in study populations and/or differences in interventions that limit HTE generalizability.