| Sumario: | Background and Objective: Class imbalance in stroke datasets often results in increased misdiagnosis rates, biased risk factor analysis, and limited clinical utility of machine learning models. This study proposed a method combining Generative Adversarial Networks (GAN) with a hard voting ensemble to enhance the accuracy of stroke risk identification, thereby supporting improved stroke prevention and control. Methods: Data were from 4,238 participants in the NHANES database (2007–2018). A balanced subset resampling strategy combined with LASSO-logistic regression was employed to identify stroke risk factors. GAN was used to generate minority class samples to achieve dataset balance, and its performance was compared with traditional data balancing methods. Four base classifiers were constructed using selected features, and predictions were integrated via hard voting, including Logistic Regression (LR), Random Forest (RF), XGBoost, and Gradient Boosting Decision Tree (GBDT). Performance was assessed using sensitivity, F1 score, and G-mean. Results: Age (OR = 1.050, P < 0.001) and depression score (DPQ-9, OR = 1.065, P < 0.001) significantly increased stroke risk. Family income-to-poverty ratio (PIR, OR = 0.911, P = 0.026) and total polyunsaturated fatty acid intake (PUFA, OR = 0.988, P = 0.046) exhibited protective effects. GAN-based data balancing improved XGBoost's sensitivity from 0.018 to 0.915. The voting ensemble achieved an F1 score of 0.956 and specificity of 1.000. SHapley Additive exPlanations (SHAP) analysis confirmed the dominant role of these factors. Conclusion: This study integrated GAN-based data generation with a hard voting mechanism to effectively address class imbalance in stroke prediction. It significantly improved model stability and clinical identification capability, providing a reliable tool to screen high-risk populations and formulate targeted prevention strategies.
|
|---|