Recurrent prediction within 1, 3, and 5 years after acute ischemic stroke based on machine learning using 10 years J-ASPECT study

Background and Purpose: We aimed to predict recurrent ischemic stroke after the first acute ischemic stroke within 1, 3, and 5 years. Methods: The subjects were patients admitted to participating J-ASPECT Study facilities with acute ischemic stroke from 2010 to 2019. A machine learning model was constructed using 117 clinical information variables to compare the predictive accuracy of the Stroke Prognosis Instrument II (SPI-II) and the Essen Stroke Risk Score (ESRS). Results: The mean area under the receiver operating characteristic curve (ROC AUC) for machine learning prediction of stroke recurrence within 1, 3, and 5 years was 0.62, 0.63, and 0.63, respectively. In contrast, the ROC AUCs for SPI-II were 0.54, 0.54, and 0.54, and ESRS were 0.55, 0.54, and 0.53 within 1, 3, and 5 years, respectively. Conclusion: The machine learning prediction model achieved better performance than classical risk scores for recurrent ischemic stroke prediction.