CDF2S: Improving stroke prediction with cluster-based under sampling and interpretable deep forest model

Abstract Stroke is a series of symptoms caused by the occlusion or hemorrhage of blood vessels supplying the brain. Cells in the affected central area of the brain can die within minutes, leading to irreversible damage. Currently, the optimal solution is early prevention based on relevant risk factors. However, imbalances in medical data can cause the performance of predictive models to degrade. In recent years, research on imbalanced data issues has mainly focused on undersampling the majority class while neglecting the minority class. In this paper, we propose a novel stroke prediction method, CDF2S, which combines a cluster-based undersampling algorithm (CBUC) and a deep forest model. CDF2S utilizes the CBUC algorithm to cluster the minority class into K clusters and splits the samples within each cluster into training and testing sets to enhance the diversity and richness of the training samples. Based on the results of 10-fold cross-validation, CDF2S outperforms state-of-the-art methods, with core metrics of accuracy, Gmean, and AUC improving by 6, 4, and 4 percentage points, respectively. Additionally, we calculate the Gini impurity of input features using the mean decrease impurity (MDI) method and conduct an interpretability analysis of the model. Detailed data and code can be found at: https://github.com/hvskghdghsjsv/CDF2S-celebral-stroke-prediction