Data-efficient machine learning approach for predicting asthma attack risk.

Area of Science:

• Respiratory Medicine
• Data Science
• Computational Biology

Background:

• Asthma is a significant global health issue, causing over 436,000 deaths in 2021.
• Accurate asthma attack prediction can improve patient outcomes and reduce healthcare costs.
• Current prediction models often require extensive data, limiting their use in data-scarce environments.

Purpose of the Study:

• To develop a data-efficient approach for predicting asthma attack risk.
• To reduce the complexity and data requirements of asthma prediction models.
• To enhance the clinical applicability of asthma risk prediction tools.

Main Methods:

• Utilized the CRISP-DM methodology for model development.
• Explored combinations of feature selection strategies, machine learning algorithms (e.g., Logistic Regression, XGBoost), and data imbalance handling techniques.
• Constructed and evaluated 120 distinct models, focusing on accuracy and efficiency.

Main Results:

• The combination of Logistic Regression (LR) and XGBoost (XGB) models with under-sampling techniques demonstrated superior performance.
• Feature selection based on XGBoost with Random Under Sampling (XGB-RUS) identified key predictors.
• Simplified feature sets, including "asthma attacks in the past year" and "SABA_ICS ratio," achieved notable predictive accuracy.
• FS1 achieved the best performance, while FS6 with only two features showed reasonable results.

Conclusions:

• Data-efficient machine learning models can effectively predict asthma attack risk.
• Simplified feature sets can maintain high accuracy, making prediction models more practical for clinical use.
• This research provides a framework for developing and validating asthma prediction models in data-limited settings.