Calibrated Probability

The foundational property where a model's predicted probability matches the observed frequency of correct predictions in the real world. If a calibrated model assigns a 70% confidence score to 100 clinical extractions, exactly 70 of those extractions should be correct.

• Reliability Diagram: A graphical tool plotting predicted probability against observed frequency, where perfect calibration forms a diagonal line
• Miscalibration Patterns: Underconfident models predict lower probabilities than actual accuracy; overconfident models predict higher probabilities than warranted
• Clinical Impact: Ensures that a 95% confidence threshold for straight-through processing truly reflects a 5% error rate, not a hidden 20% failure rate

The primary quantitative metric measuring the weighted average of the difference between predicted confidence and actual accuracy across all probability bins. ECE partitions predictions into buckets and computes the absolute gap between each bucket's average confidence and its empirical accuracy.

• Calculation: ECE = Σ (|B_m|/n) × |acc(B_m) − conf(B_m)|, where B_m represents each bin
• Lower is Better: An ECE of 0.02 indicates only a 2% average deviation from perfect calibration
• Clinical Thresholding: ECE directly informs how reliably a confidence threshold of 0.90 can segregate high-risk cases requiring human review from low-risk cases eligible for automation

A critical distinction between two desirable but often conflicting properties. Sharpness measures how close predictions are to 0 or 1, reflecting decisiveness. Calibration measures how well those predictions reflect true likelihood.

• Overconfident Models: High sharpness but poor calibration—predictions cluster near 0 and 1 but are frequently wrong
• Underconfident Models: Low sharpness with predictions hovering near 0.5, even when the model is often correct
• Clinical Workflow Design: A well-calibrated but less sharp model may route more cases to human review, while a sharp but miscalibrated model risks dangerous automation of incorrect extractions

A proper scoring rule that jointly evaluates both calibration and refinement, decomposing overall predictive performance into distinct components. The Brier Score is the mean squared error between predicted probabilities and binary outcomes.

• Decomposition: Brier Score = Calibration Loss + Refinement Loss − Uncertainty
• Calibration Loss: Measures the reliability component—how well predicted probabilities match observed frequencies
• Refinement Loss: Measures resolution—how well the model separates positive from negative cases, independent of calibration
• Clinical Benchmarking: A Brier Score below 0.01 indicates excellent overall probabilistic performance suitable for high-stakes medical decision support

The property that calibration remains stable when the underlying data distribution changes between training and deployment. A model calibrated on one hospital system's records may become miscalibrated when applied to a different patient population.

• Covariate Shift: Changes in input feature distributions that can degrade calibration without affecting discrimination
• Label Shift: Changes in the prevalence of clinical conditions that directly distort probability estimates
• Recalibration Cadence: Clinical AI systems require periodic recalibration using recent production data to maintain reliability as documentation practices and patient demographics evolve

A predefined probability score below which a model's prediction is flagged for manual review. Calibrated probability ensures this threshold is semantically meaningful—a threshold of 0.9 genuinely corresponds to a 90% empirical likelihood of correctness.

• Directly controls the Straight-Through Processing (STP) Rate
• Setting thresholds without calibration leads to unpredictable review queue volumes
• Typical clinical settings use tiered thresholds: auto-approve (>0.95), review (0.5-0.95), auto-reject (<0.5)

A structured classification of model failure modes used by reviewers to tag corrections. Calibrated probability reveals which error types correlate with specific confidence bands.

• Low-confidence errors often indicate out-of-distribution inputs
• High-confidence errors signal systematic model blind spots requiring retraining
• Enables targeted active learning by identifying the error categories where human feedback is most valuable

The gradual deviation of a human annotator's judgment from the annotation guideline over time. Calibrated probability serves as an anchor—when a reviewer consistently overrides high-confidence predictions, it may indicate drift rather than model error.

• Compare reviewer override rates against model confidence bands
• Flag reviewers with anomalous disagreement patterns on high-calibration predictions
• Use as input for Inter-Annotator Agreement (IAA) recalibration sessions

The percentage of clinical documents processed entirely by AI without human intervention. Calibrated probability makes STP rate a reliable metric—you can mathematically bound the expected error rate of auto-processed cases.

• STP Rate = proportion of predictions above the confidence threshold
• With calibration, a 95% threshold guarantees ≤5% expected error in auto-processed cases
• Trade-off analysis: increasing STP by 10% may increase clinical risk by a quantifiable margin