Using machine learning models to improve stroke risk level classification methods of China national stroke screening

Area of Science:

• Public Health
• Medical Informatics
• Machine Learning

Background:

• Stroke poses a significant health burden in China due to high incidence, prevalence, and mortality.
• The national stroke screening program identifies high-risk individuals (aged 40+) based on factors like hypertension, diabetes, and smoking.
• Current screening limitations arise from inability to classify risk with missing data, impacting intervention efficiency and statistical accuracy.

Purpose of the Study:

• To develop and evaluate machine learning models for improved stroke risk classification.
• To address the challenge of missing risk factor data in stroke screening.
• To enhance the efficiency and accuracy of national stroke interventions.

Main Methods:

• Utilized 2017 national stroke screening data.
• Processed imbalanced training data using oversampling and undersampling techniques.
• Developed and compared various machine learning models including logistic regression, Naïve Bayesian, Bayesian network, decision tree, neural network, random forest, bagged decision tree, voting, and boosting models.

Main Results:

• The boosting model with decision trees achieved the highest recall (99.94%).
• The random forest model demonstrated the highest precision (97.33%).
• The random forest model improved recall by approximately 2.8% compared to the existing method, enabling identification of thousands more high-risk individuals annually.

Conclusions:

• Machine learning models offer a superior alternative to current stroke risk screening methods.
• These models effectively handle missing data, reduce unnecessary rescreening, and optimize intervention expenditures.
• The national program can adopt these classification models to enhance screening effectiveness based on practical needs.