Comparison of outcome prediction models post-stroke for a population-based registry with clinical variables collected at admission

Aim: The ability to predict outcomes can help clinicians to better triage and treat stroke patients. We aimed to build prediction models using clinical data at admission and discharge to assess predictors highly relevant to stroke outcomes. Methods: A total of 37,094 patients from the Taiwan Stroke Registry (TSR) were enrolled to ascertain clinical variables and predict their mRS outcomes at 90 days. The performances (i.e., the area under the curves (AUCs)) of these independent predictors identified by logistic regression (LR) based on clinical variables were compared. Results: Several outcome prediction models based on different patient subgroups were evaluated, and their AUCs based on all clinical variables at admission and discharge were 0.85-0.88 and 0.92-0.96, respectively. After feature selections, the input features decreased from 140 to 2-18 (including age of onset and NIHSS at admission) and from 262 to 2-8 (including NIHSS at discharge and mRS at discharge) at admission and discharge, respectively. With only a few selected key clinical features, our models can provide better performance than those previously reported in the literature. Conclusion: This study proposed high performance prognostics outcome prediction models derived from a population-based nationwide stroke registry even with reduced LR-selected clinical features. These key clinical features can help physicians to better focus on stroke patients to triage for best outcome in acute settings.