Clinical ML: Why Interpretability Often Beats Accuracy

From the Heart Failure Readmission project

Context

In healthcare ML, the most accurate model is not always the most valuable. During my heart failure readmission work using the MIMIC-IV clinical database, I had to choose between a higher-scoring black-box model and a more interpretable, clinically-actionable model. I selected Logistic Regression over XGBoost, despite a 6% accuracy gap, because interpretability, clinical utility, and regulatory compliance (HIPAA/FDA guidance) dominated.

Decision Framework

Weighted scoring favored Logistic Regression for clinical deployment, even though XGBoost led on pure metrics.

Technical Highlights

Outcomes

Takeaways


Read the full project and implementation details in the repository: heart_failure_readmission.

Deutsche Version