Data augmentation alters feature importance in XGBoost for CVD prediction.

Area of Science:

• Cardiovascular Disease Research
• Machine Learning in Medicine
• Data Science

Background:

• Machine learning models are vital for cardiovascular disease (CVD) prediction.
• Dataset size and class imbalance often limit model performance.
• The impact of data augmentation on model interpretability and feature importance is poorly understood.

Purpose of the Study:

• To investigate how data augmentation strategies affect the performance and feature importance of Extreme Gradient Boosting (XGBoost) models for CVD prediction.
• To compare the effects of Synthetic Minority Over-sampling Technique (SMOTE) and Wasserstein Generative Adversarial Network with Gradient Penalty (WGAN-GP) augmentation.

Main Methods:

• An ablation study was conducted using a public CVD dataset.
• Three XGBoost models were compared: baseline (original data), SMOTE-augmented, and WGAN-GP-augmented.
• Model performance was evaluated using accuracy, F1-score, and AUC; feature importance was assessed using the Gain metric.

Main Results:

• All models showed high predictive performance; the SMOTE-augmented model achieved 1.0 accuracy and AUC.
• Data augmentation significantly altered feature importance rankings.
• 'slope' became the dominant feature in augmented models (SMOTE: Gain 27.49, WGAN-GP: Gain 36.68), unlike the baseline model (Gain 7.01).

Conclusions:

• Data augmentation can reshape the predictive strategy of machine learning models.
• For high-quality datasets, augmentation may re-prioritize features rather than solely improving accuracy.
• Evaluating synthetic data's impact on model interpretability is crucial before clinical application.