Fig. 1
Overview of PAI and BRL modeling workflow. a The first step is PAI regression modeling which takes in data listed in Table 1 and trains on the whole data set to predict both actual and counterfactual post-treatment scores for individual patients (actual scores can be compared with predicted actual scores and resulted in an R2 = 0.58 as shown in Fig. 2). b The PAI Thresholding step thresholds the difference between actual and counterfactuals to create indication labels for each patient. A treatment-indicated subgroup had a treatment effect size of Cohen’s d = 1.51 as an intermediate assessment, but explainability of model decisions needs improvement, so the BRL step addresses this need. c The BRL modeling uses fivefold cross-validation to assess generalization ability to unseen samples. The predictions generated for test samples over all folds had an accuracy of 74.1% and an AUC of 0.74 for this classifier. Importantly, it emits an explainable rule list for each fold. d The final step is assessing the treatment effect of the treatment-indicated subgroup identified by BRL (Cohen’s d = 1.24 as seen in Fig. 5)