Tackling Imbalanced Data in Chronic Obstructive Pulmonary Disease Diagnosis: An Ensemble Learning Approach with...

Area of Science:

• Medical Informatics
• Artificial Intelligence in Healthcare
• Pulmonology

Background:

• Chronic Obstructive Pulmonary Disease (COPD) poses a significant global health challenge, with Taiwan experiencing rising prevalence and mortality.
• 14-day readmission is a critical indicator of COPD disease instability and transitional care effectiveness, influenced by patient vulnerability and healthcare system factors.

Purpose of the Study:

• To develop an AI-based prediction model for forecasting 14-day readmission risk in COPD patients.
• To address the challenge of class imbalance in medical data, common in machine learning models for predicting rare events.

Main Methods:

• Utilized real-world clinical data from Kaohsiung City hospitals.
• Employed data augmentation techniques, specifically CTGAN and Kernel Density Estimation (KDE), to address class imbalance.
• Constructed a stacking ensemble model with six base learners, comparing its performance against baseline models (XGBoost, AdaBoost, Random Forest, LightGBM) across various configurations and recall thresholds.

Main Results:

• Kernel Density Estimation (KDE) combined with stacking demonstrated superior and stable performance, particularly under high-recall targets, compared to baseline models.
• Ablation experiments revealed that the K-Nearest Neighbors (KNN) base model was crucial for the stacking classifier's performance, especially with KDE-generated data.
• The final recommended model configuration, integrating KDE and stacking, significantly improved precision, F1-score, and specificity.

Conclusions:

• The developed AI framework, leveraging KDE data augmentation and stacking ensemble learning, offers a robust solution for predicting COPD 14-day readmissions.
• This predictive model can facilitate early identification of high-risk patients, enabling timely interventions to reduce avoidable readmissions and enhance patient care quality.