Enhancing Clinical Prediction Through Feature Engineering

The use of raw, high-dimensional tabular data is central to current clinical predictive modeling. While this data is informative, the way it is structured and represented for machine learning models plays a critical role in both predictive performance and interpretability. Raw clinical variables often lack contextual aggregation and methodological framing, which can limit model generalization and obscure the reasoning behind predictions.

Using hospital and ICU data from MIMIC-IV, we construct and evaluate engineered features derived from patient stays. We assess their impact on predictive performance, robustness, and interpretability, and analyze which feature transformations contribute most significantly to model behavior. Whereas many prior studies emphasize maximizing predictive metrics alone, this work prioritizes having interpretable features and methodological rigor to ensure that model predictions align with clinically meaningful patterns. By examining not only how well models perform but also why they make specific predictions, this work aims to provide clearer methodological guidance for clinical machine learning applications.