Functions for assessing and improving probability calibration of classification models. Well-calibrated probabilities are essential for clinical decision-making.
Details
## Why Calibration Matters
A model with high AUC can still produce poorly calibrated probabilities. For clinical use, we need: - When model says 80 - Brier score decomposes into discrimination + calibration + uncertainty - ECE (Expected Calibration Error) measures mean deviation from perfect calibration
## Calibration Repair
Post-hoc calibration can improve probability estimates: - **Platt scaling**: Fits logistic regression on predictions - **Isotonic regression**: Non-parametric monotonic recalibration - **Temperature scaling**: Single parameter softmax recalibration