Enhancing Clinical Predictive Modeling through Model Complexity-Driven Class Proportion Tuning for Class Imbalanced...

Area of Science:

• Medical Informatics
• Machine Learning
• Public Health

Background:

• Class imbalance is a common problem in medical predictive modeling.
• Existing methods for class imbalance often overlook model-specific characteristics.

Purpose of the Study:

• To propose a new method for addressing class imbalance in clinical predictive models.
• To demonstrate that optimal class proportions are dependent on model complexity.

Main Methods:

• Developed a novel approach to determine class proportions based on model complexity.
• Applied and validated the method using the opioid overdose prediction problem.
• Conducted rigorous regression analysis to confirm the theoretical framework.

Main Results:

• The proposed method achieved significant performance gains in opioid overdose prediction.
• Demonstrated a statistically significant correlation between model complexity hyperparameters and optimal class proportions.
• Showcased the effectiveness of individualized tuning of class proportions for specific models.

Conclusions:

• The optimal class proportions for predictive models are intrinsically linked to their complexity.
• This model-complexity-aware approach offers improved performance over traditional methods for imbalanced datasets.
• The findings provide a robust theoretical framework for optimizing predictive models in healthcare.