Abstract Background and aims Rescue intracranial stenting after failed mechanical thrombectomy remains a clinical dilemma, with observational data suggesting benefit but significant procedural risks and lifelong antiplatelet requirements. Current practice lacks tools to identify which patients are most likely to benefit. We developed a causal machine learning framework to estimate individualized treatment effects for rescue stenting decisions. Methods We applied T-Learner methodology with stacking ensemble models to the SAINT multicenter cohort (n=393; 195 rescue stenting, 198 failed MT only; baseline mRS 0-2). Separate models predicted outcomes under each treatment scenario (μ0: failed MT, μ1: rescue stenting), with Conditional Average Treatment Effect (CATE) quantifying individual-level benefit. Outcomes included functional independence (mRS 0-2), favorable outcome (mRS 0-3), symptomatic ICH, and 90-day mortality. Internal validation used Harrell’s optimism-corrected bootstrap (5000 iterations). A ≥15% absolute benefit threshold (NNT≤7) defined clinically meaningful effect given procedural risks. Results Optimism-corrected AUCs were excellent: functional independence 0.90(95%CI:0.77-0.99) for failed MT and 0.89(0.84-0.93) for rescue stenting models. Mean CATE for functional independence was +29.0 percentage points, with 91.7% of patients showing predicted benefit. However, only 75% exceeded the ≥15% clinically meaningful threshold. Patients with shorter time-to-puncture(importance:0.26), younger age(0.25), higher ASPECTS(0.16), fewer thrombectomy passes(0.15), and moderate NIHSS(0.13) demonstrated greatest predicted benefit. Mortality reduction paralleled functional gains(+14.3%). Conclusions Causal machine learning enables individualized rescue stenting. Approximately one-quarter of patients do not meet the threshold for clinically meaningful benefit. This approach may inform both clinical decision-making and future trial designs requiring prospective validation. Conflict of interest None
Doheim et al. (Fri,) studied this question.