Fig. 4: Treatment effect estimation and discovering heterogeneous subgroups.

a We plot the predictions from baseline of a single patient’s serum immunoglobulins over both the factual treatment (here IRd - top) and the counterfactual treatment (Rd - bottom), demonstrating SCOPE's ability to model counterfactuals. Dots represent the observed values, and solid lines represent the predictions of SCOPE from baseline. The variables with units g/L are measured along the right axis (SPEP Kappa, Lambda, and Monoclonal Protein), and variables with units mg/L are measured along the left axis (IgA and IgG). b Diagram of a proof-of-concept subgroup discovery analysis. c On the left, we show the Kaplan-Meier curves for the original treatment and control groups in a left-out set of MM2 patients. In the middle, we show the Kaplan-Meier curves for the treatment and control groups in a learned subgroup of MM2 that suggest greater differential survival between IRd and Rd for this patient subgroup. On the right, we show the Kaplan-Meier curves for the interpretable subgroup learned to replicate the subgroup found with our model. The p-values were computed with a log-rank test. d Visualization of the learned decision tree, which was trained using the subgroup assignments as labels and patient baseline data as features. Each node contains the number of patients, and the proportion of the samples who are labeled as getting Rd (P₀) or IRd (P₁), respectively. A node is colored red if P₀ > P₁, green if P₀ < P₁, and blue if P₀ ≈ P₁.