Optimizing remote mechanical thrombectomy training using causal machine learning: a multicenter study

Remote, live-streamed mechanical thrombectomy (MT) training is a novel educational strategy, but its impact on patient outcomes remains unclear. The authors aimed to move beyond average treatment effects to identify which clinician subgroups derive the most significant clinical gains from cumulative training exposure.

In this prospective multicenter cohort study, 1010 clinicians participated in standardized, live-streamed MT training. The authors applied causal inference techniques, including double/debiased machine learning (DML), causal forest analysis, and Bayesian mixed-effects models, to evaluate the impact of cumulative training hours on 90-day patient outcomes. A Bayesian generalized linear mixed model was used to map personalized dose-response trajectories stratified by baseline clinical competency and surgical experience.

Longer cumulative training hours was associated with improved procedural performance and better 90-day functional outcomes. DML models consistently showed significant treatment effects (0.83-0.85, p < 0.001). Bayesian mixed-effects modeling confirmed a strong association between training duration (OR 2.2, 95% CI 1.58-3.02) and clinician competency (score range 0-100; OR 3.7, 95% CI 2.79-4.86). Notably, while the training conferred universal clinical benefits across all baseline levels, the learning curve demonstrated continuous, linear improvements that were particularly pronounced among high-competency practitioners.

Structured remote MT training was linked to improved procedural efficiency and functional recovery after stroke. These findings support the use of data-driven, personalized, and experience-stratified training strategies to enhance clinical practice and patient outcomes in stroke care.