Abstract WP94: Forecasting Outcomes After Mechanical Thrombectomy: Predictive Modeling of Neurological Improvement in the SELECT Database

Abstract

Listen

Translate article

Introduction: Outcome prediction and prognostication are at the heart of determining next steps in stroke care. We aim to evaluate two variable selection strategies to provide optimal prediction of functional independence among post-EVT patients at admission, day 1, and time of discharge using data from the SELECT prospective cohort study. Methods: Prediction strategies were created using baseline demographics, clinical presentation and imaging characteristics for the admission model, procedure characteristics and outcomes, clinical status, and baseline factors for the day 1 model, and follow-up infarct volume, cerebral edema, clinical status, and previously mentioned factors for the discharge model. Model selection used backwards stepwise regression and least absolute shrinkage and selection operator (LASSO) strategy. Patients with unavailable data were excluded. Calibration and internal validation was performed using bootstrapping for 50 repetitions. Results: 149 EVT patients with all available data points were included. Functional independence (mRS 0-2) was observed in 89 (60%) patients. LASSO and stepwise selection methods selected different variables at baseline, with LASSO providing reasonably better model performance over stepwise model selection. Both stepwise and LASSO regression methods selected baseline ASPECTS and NIHSS at day 1 for day 1 model, and baseline ASPECTS, NIHSS at day 1 and NIHSS at day 7 for discharge model. Day 1 and discharge models demonstrated significantly better predictive performance in bootstrap validation over baseline models [Table]. Conclusions: Clinical symptoms in the post-acute phase were the strongest predictors of functional independence at 90-day follow-up and were selected over procedural factors or cerebral edema. Prediction of functional independence at baseline involved several clinical and imaging variables with lower predictive performance than post-acute phase models.