The machine learning model classified acute ischemic stroke onset ≤4.5 hours with 94% AUC and 96% PPV, supporting time-based treatment decisions for unwitnessed cases.
Does an MRI-informed machine learning model accurately classify acute ischemic stroke onset as ≤4.5 hours versus >4.5 hours?
An MRI-informed machine learning model can classify early versus late acute ischemic stroke onset with high specificity and positive predictive value, potentially supporting time-based treatment decisions for unwitnessed strokes.
Time since last known well (LKW) is a key factor in determining eligibility for reperfusion therapies in acute ischemic stroke (AIS) patients. However, many patients present to the emergency room with unwitnessed and therefore unknown onset time of cerebral ischemia. This study aimed to develop a machine learning model to classify early and late presenting strokes with respect to the treatment window for thrombolytic therapy (≤4.5 hours). A retrospective analysis was performed on a prospectively collected cohort of N = 350 AIS patients with witnessed stroke onset who underwent MRI imaging at Ascension Texas hospitals from March 2018 to November 2024, with 54 classified as late strokes. An extreme gradient boosting (XGBoost) classification model was trained using baseline covariates (age, sex, modified Rankin Scale, NIH Stroke Scale, vascular risk factors), RAPID AI perfusion metrics including hypoperfusion intensity ratio (HIR), and quantified T2-FLAIR signal intensity. Model performance was evaluated on held-out data during stratified 10-fold cross validation. The model was tuned to maximize specificity by minimizing false positives (late strokes misclassified as early) while targeting sensitivity >80% to limit false negatives (early strokes misclassified as late). In validation data, the model classified onset ≤4.5 hours versus >4.5 hours with an area under the precision-recall curve of 94% (95% CI: 90 – 98), 96% positive predictive value (93 – 98), 85% specificity (73 – 93), and 72% sensitivity (67 – 77). At our cohort’s late-stroke prevalence (15%), the model yields approximately 2 false positives and 13 true negatives per 100 patients (and 61 true positives, 24 false negatives). Low sensitivity, defined as missed treatment opportunities, was the trade-off for minimizing false positives. T2-FLAIR signal intensity ratio (SIR), NIH Stroke Scale, and HIR were the strongest contributors to model predictions. Re-training the model defining the treatment window as <6.0 hours achieved similar results. Our MRI-informed model classified early versus late AIS onset with high specificity and predictive value. This classification model incorporates imaging and clinical features to support time-based treatment decisions in patients with wake-up or unwitnessed stroke onset.
Ramirez et al. (2026) studied this question. The machine learning model classified acute ischemic stroke onset ≤4.5 hours with 94% AUC and 96% PPV, supporting time-based treatment decisions for unwitnessed cases.