Endovascular Thrombectomy for Minor Ischemic Stroke: A Machine Learning Study of Treatment Effect Heterogeneity.

Background

AND

Objectives

Endovascular thrombectomy (EVT) is standard of care for large vessel occlusion (LVO) stroke with moderate-to-severe deficits, but its role in minor stroke (NIHSS <6) remains uncertain. We characterized EVT's association with outcomes, examined complications as potential mediators of that association, and explored heterogeneity of treatment effect (HTE).

Methods

Retrospective cohort study using the Nationwide Readmissions Database (2016-2022) in adults with non-elective LVO stroke and NIHSS <6. Primary outcomes were routine discharge and in-hospital mortality; secondary outcomes included intracranial hemorrhage (ICH), cerebral edema, and non-neurological complications. Propensity score matching (PSM), causal mediation analysis, and a causal forest model estimated treatment effects, mediated harm, and HTE, respectively.

Results

Among 237,812 weighted admissions (14,440 EVT; 223,372 no EVT), PSM showed no difference in routine discharge (49.4% vs. 50.3%; P=0.484) but higher mortality with EVT (4.6% vs. 1.9%; P<0.001). EVT was associated with higher ICH (15.0% vs. 7.3%), cerebral edema (6.2% vs. 3.8%), and respiratory complications (10.7% vs. 5.7%; all P<0.001). ICH was the largest mediator of EVT-associated mortality (19.4%), followed by respiratory complications (17.2%) and cerebral edema (10.2%). Causal forest identified significant HTE (P<0.001); higher NIHSS, absence of thrombolysis, older age, and facial weakness were the strongest predictors of predicted benefit. A favorable subgroup (31.2%) showed higher routine discharge with EVT (46.3% vs. 43.7%; P=0.023), though mortality remained elevated across all groups.

Conclusions

EVT for minor LVO stroke was associated with higher mortality without improved routine discharge. Because some patients may be selected for EVT after deterioration, these are associations rather than causal effects. Individualized selection and randomized data are needed.