Explainable AI for Data-Driven Design of High-Dimensional Predictive Studies

Explainable AI can improve interpretable models without sacrificing transparency: The framework enhanced a traditional statistical model (Cox Proportional Hazards) by using AI to identify useful features and interactions, improving predictive performance (C-index from 0.805 to 0.815) while maintaining clinical interpretability—addressing the common trade-off between accuracy and explainability.

AI recommendations can simplify complex model design: When faced with hundreds of potential features, the method systematically recommended excluding 23 features and adding 221 feature interactions, demonstrating how AI can guide clinicians through otherwise overwhelming design choices in a data-driven way.

Maintaining model calibration is crucial for clinical trust: Beyond improving discrimination (C-index), the method preserved calibration metrics (intercept and slope remained near optimal values), which is essential for clinicians to trust predicted probabilities in real decision-making scenarios.

This approach bridges the "black box" problem in healthcare: By using explainable AI techniques to translate complex patterns into understandable recommendations (feature exclusion, non-linear terms, interactions), the framework enables high-performing predictions while keeping the underlying model transparent enough for clinical validation and trust.

Real-world validation matters: The evaluation used a large dataset of 245,614 patients predicting fall injuries, suggesting the method's practical applicability to actual clinical problems rather than theoretical scenarios.