Solving the class imbalance problem using ensemble algorithm: application of screening for aortic dissection.

Area of Science:

• Medical Informatics
• Machine Learning
• Cardiovascular Disease Research

Background:

• Class imbalance, an imbalance between positive and negative outcomes, is a common challenge in medical datasets.
• This imbalance significantly complicates the development of accurate predictive models, particularly for rare diseases like aortic dissection (AD).

Purpose of the Study:

• To develop and validate an effective integrated approach to address class imbalance in medical data.
• To create a practical early screening model for the rare cardiovascular disease, aortic dissection (AD).

Main Methods:

• Combined data-level methods, cost-sensitive learning, and bagging.
• Applied feature selection using statistical analysis and logistic regression.
• Integrated Support Vector Machine (SVM) weak classifiers with undersampling and bagging for a strong classifier, validated on 523,213 patient records with a 1:65 AD to non-AD ratio.

Main Results:

• The ensemble model achieved a sensitivity of 82.8% and specificity of 71.9% with a training time of 56.4 seconds.
• Demonstrated low sensitivity variance (19.58 × 10^-3) across seven-fold cross-validation.
• Outperformed common ensemble algorithms (AdaBoost, EasyEnsemble, Random Forest) and single machine learning methods (logistic regression, decision tree, KNN, BP, SVM).

Conclusions:

• The integration of feature selection, undersampling, cost-sensitive learning, and bagging effectively overcomes class imbalance in medical datasets.
• A practical screening model for aortic dissection (AD) was developed, providing decision support for early-stage detection.
• This approach holds potential for improving the screening of rare diseases with imbalanced data.