Abstract / Summary
Abstract Background Accurate estimation of the individualized treatment effect (ITE) is critical for supporting personalized decision-making in acute ischaemic stroke, while the average treatment effect (ATE) provides a cohort-level benchmark for method comparison and context. However, the consistency and reliability of different causal inference methods for this purpose remain uncertain. Methods A total of 12,045 ischaemic stroke patients registered in the multicentre German Stroke Registry Endovascular Treatment were eligible for the study. Multiple causal inference models including Causal Forest, S-Learner, T-Learner, X-Learner, and ensemble approaches were compared for individualized and average treatment effect estimation in accordance with the Predictive Approaches to Treatment effect Heterogeneity statement. ITE was defined as the patient-specific difference in predicted risk of 90-day poor functional outcome between the counterfactual states of achieving successful recanalization versus not achieving successful recanalization after mechanical thrombectomy, and the ATE was defined as the cohort mean of these individual differences. Model performance was assessed using AUC, ECE, and Brier score, and the agreement of individual treatment rules was assessed using pairwise comparisons. Multivariable regression analysis identified variables associated with negative estimated ITE. Results Across six ML approaches, recommendations based on the estimated benefit of successful recanalization after mechanical thrombectomy were generally consistent. Pairwise ITR agreement ranged from 90.1% to 96.4%. Models estimated benefit from successful recanalization for most patients (90.4 to 97.3%; median: 93.6%). ATE estimates ranged from 18.3% to 25.6%. Estimated ITEs correlated moderate to strong across models (0.58 to 0.92; median: 0.76). Multivariable analyses showed that poorer pre-stroke functional status, higher NIHSS at admission, and older age were consistently associated with negative estimated ITE across models. Conclusions Machine Learning-based causal inference models provided generally consistent estimates of individualized and average effects associated with successful recanalization after mechanical thrombectomy. Percentage agreement in treatment rules was high, largely reflecting the high overall treatment rate, but beyond-chance agreement was only slight-to-fair (median Cohen’s kappa = 0.32; range 0.20–0.58), indicating that individual-level recommendations can be sensitive to modelling choices.