Revisiting the High-Benefit Patient: Generic Machine Learning Inference for Intensive Blood Pressure Control

We applied the model-free GenericML framework using robust estimators (the Best Linear Predictor, Group Average Treatment Effects, and baseline profile classification) to reassess treatment effect heterogeneity (HTE) of intensive blood pressure (BP) using pooled data from two large-scale randomized control trials on BP controls. Among 10,712 participants with up to three years of follow-up, we estimated the differential risks of the primary cardiovascular outcome. Intensive BP control produced a modest average risk reduction (BLP β ₁ = −0.0136), but evidence for HTE was not statistically significant (BLP β ₂ p = 0.443; GATES top–bottom contrast p = 0.471). Although CLAN grouped participants with metabolically adverse profiles into the highest predicted-benefit stratum, their treatment response did not differ significantly from that of the lowest-benefit group. Overall, while machine learning can identify clusters of high-risk baseline phenotypes, group-level inference revealed no meaningful HTE, underscoring the need for caution when interpreting individualized treatment predictions.