Explainable artificial intelligence-driven ensemble learning for asthma risk prediction using machine and deep...

Area of Science:

• Pulmonary Medicine
• Computational Biology
• Data Science

Background:

• Asthma is a chronic respiratory condition affecting millions globally, leading to significant healthcare costs and reduced quality of life.
• Early prediction and management of asthma risk are crucial for preventing exacerbations and improving patient outcomes.
• Current prediction methods may lack the accuracy and real-time applicability needed for effective clinical intervention.

Purpose of the Study:

• To develop a comprehensive asthma risk prediction model utilizing advanced machine learning and deep learning techniques.
• To enhance model accuracy and interpretability for clinical decision support.
• To create a real-time, user-friendly interface for immediate asthma risk assessment.

Main Methods:

• Employed Recursive Feature Elimination and Extra Trees Classifier for optimal feature selection.
• Utilized Synthetic Minority Over-sampling Technique (SMOTE) to address class imbalance in the dataset.
• Optimized 12 machine learning and deep learning models, including XGBoost, Random Forest, SVM, MLP, CNN, RNN, and ANN, through hyperparameter tuning.

Main Results:

• An ensemble approach combining XGBoost and CatBoost with soft voting achieved the highest prediction accuracy of 93.61% after hyperparameter tuning.
• Interpretable AI methods (SHAP and LIME) were used to explain model predictions and identify key contributing factors.
• A functional Flask server and web interface were deployed for real-time asthma risk prediction.

Conclusions:

• The developed framework provides an accurate and explainable asthma risk prediction tool.
• The high accuracy (93.61%) and real-time capabilities demonstrate significant clinical applicability.
• This approach has the potential to improve early intervention strategies and patient management for asthma.