Initial Diagnosis of Stroke Based on SMOTE- ENN and Hybrid Machine Learning

In this paper, a stroke screening model based on SMOTE-ENN and hybrid machine learning is constructed. In view of the imbalance of medical data categories and the fact that the cost of missed diagnosis is significantly higher than that of misdiagnosis in the primary screening scene of stroke, this study constructs a multi-dimensional optimization framework : in the data layer, firstly, the quality of the data set is improved by preprocessing technology, and secondly, SMOTE-ENN is used to balance the data distribution; then, in the model layer, the integration strategies of Bagging, Boosting, Stacking and Voting are integrated. At the same time, a cost-sensitive learning mechanism is embedded in the parameter layer; implement a dynamic threshold optimization strategy at the decision-making layer. Experiments show that Soft Voting ( RF : LightGBM : XGBoost = 2 : 3 : 1 ) performs best : while maintaining a high accuracy ( 70.67 % ), the false negative rate is reduced to 2.92 % ( sensitivity is as high as 97.08 % ), the false positive rate is controlled at 29.86 %, and the AUC is 0.836. The balance characteristics of the model ’ very low false negative rate-moderate false positive rate ’ make it especially suitable for primary screening in primary medical institutions and family health monitoring scenarios, providing reliable technical support for early warning of stroke.