Prediction of Stroke Hematoma Expansion Using a Machine Learning Model with Stacked Generalization

Hemorrhagic stroke is a prevalent global disease. Following hemorrhagic stroke, hematoma expansion emerges as a crucial prognostic risk factor. Currently, most prediction models for hematoma expansion rely on traditional machine learning algorithms [1]. However, these standalone machine learning algorithms often struggle to meet the desired accuracy levels for prediction. Additionally, public sample data frequently exhibit imbalanced characteristics, posing challenges during model training. To address this issue of imbalanced sample data, this study employs Synthetic Minority Over-sampling Technique (SMOTE) to oversample the minority samples within the dataset and generate more representative instances [2]. Subsequently, multiple models including Logistic Regression (LR), Support Vector Classifier (SVC), Extreme Gradient Boosting (XGB), Multi-Layer Perceptron (MLP), and Light Gradient Boosting Machine (LGBM) are trained and evaluated accordingly. Based on comprehensive evaluation results, we propose a stacked generalization-based model fusion approach that combines and stacks the aforementioned five models together. The findings demonstrate that the fused model after stacked generalization exhibits superior predictive performance by accurately forecasting hematoma expansion among patients—a significant contribution towards enhancing patient prognosis.