Prediction of the short-term prognosis of acute ischaemic stroke in patients with high treatment platelet reactivity using explainable machine learning

The aim of this study is to establish and validate an optimal explainable prediction model based on a machine learning (ML) approach to predict the short-term prognosis in high on-treatment platelet reactivity (HTPR) individuals with acute ischaemic stroke (AIS). Using individual basic characteristics, blood test indices, and the CYP2C19 genotype, a model to predict a poor functional prognosis (modified Rankin scale score ≥ 3) was constructed based on ML models, including logistic regression, support vector machine, decision tree, random forest (RF), extreme gradient boosting, and light gradient boosting machine. On this basis, global and local interpretability techniques were used to interpret selected ML models and explore the risk factors affecting the short-term prognosis of AIS in patients with HTPR. In this study, the performance of the model was futher evaluated through sensitivity analysis and subgroup analysis. A total of 515 AIS patients with HTPR were retrospectively enrolled, and approximately 129 (25%) had a poor outcome in the short term. Among the 6 ML models, RF performed best in discriminative ability in terms of area under the curve (0.84 [0.71-0.97]), accuracy (0.80 [0.71-0.89]), and precision (0.71 [0.61-0.81], which are far superior to the other models. Interpretability techniques showed that high levels of diastolic blood pressure, blood urea nitrogen, homocysteine, C-reactive protein, white blood cells, and CYP2C19 poor metabolizers were significant predictors of a poor prognosis of AIS in patients with HTPR. The risk prediction model for AIS patients with HTPR based on RF algorithms has high predictive power. By applying interpretability methods, the model's transparency and clinical usability were enhanced, offering a reference for the clinical prevention and treatment of HTPR.