Handling Class Imbalance in Machine Learning-based Prediction Models: A Case Study in Asthma Management.

Area of Science:

• Medical Informatics
• Machine Learning in Healthcare
• Pulmonology

Background:

• Asthma attack prediction tools can improve patient outcomes.
• Machine learning (ML) models for asthma prediction face challenges due to severe class imbalance.
• Effective handling of class imbalance is crucial for reliable ML models in asthma prognosis.

Purpose of the Study:

• To systematically compare class imbalance handling techniques for asthma risk prediction.
• To evaluate the performance of Synthetic Minority Oversampling Technique (SMOTE) and cost function customization.
• To identify optimal methods for improving ML-based asthma prognosis models.

Main Methods:

• Utilized data from 9,835 asthma patients from the MIMIC-IV database.
• Applied five class imbalance handling methods, including SMOTE and cost function customization.
• Developed and compared logistic regression (LR) and extreme gradient boosting (XGBoost) models for three prediction tasks with varying class imbalance ratios.

Main Results:

• Cost function customization significantly outperformed SMOTE-based methods across all prediction tasks.
• XGBoost with cost function customization achieved the highest prediction performance (AUC = 0.72) for the most imbalanced outcome.
• The study demonstrated the superiority of cost function customization over oversampling for asthma management prediction.

Conclusions:

• Cost function customization is a more effective strategy than oversampling for addressing class imbalance in asthma prediction.
• Improved ML models can lead to early asthma attack warnings, enhancing patient management and quality of life.
• This methodological advancement offers a solution for developing more accurate clinical prediction tools for asthma.