Read the ROC curve before reducing it to AUC
The ROC curve plots sensitivity against the false-positive rate as the classification threshold changes. Curves that bow toward the upper-left corner show better ranking across thresholds. A curve near the diagonal provides little discrimination. Curves can cross, so two models with similar AUC may behave differently in the region that matters.
Choose an operating region from the consequences of false positives and false negatives—not from visual closeness to the corner alone. This page interprets aggregate model output; it does not set a clinical or individual decision threshold.
Attributed AUC reference bands
| AUC | Common label | Source/context | Category |
|---|---|---|---|
| .50 | No discrimination | ROC ranking interpretation3 | FORMAL |
| .70 to <.80 | Acceptable | Hosmer–Lemeshow-style bands reproduced by Mandrekar3 | DISCIPLINE CONVENTION |
| .80 to <.90 | Excellent | Same methodological diagnostic-test context3 | DISCIPLINE CONVENTION |
| ≥.90 | Outstanding | Same methodological diagnostic-test context3 | DISCIPLINE CONVENTION |
These labels are a cited convention, not a universal performance law. An AUC of .79 and .80 are not meaningfully separated by the label boundary. Task prevalence, spectrum of cases, uncertainty, validation design, and operating costs can all change whether the result is useful. Where statistical thresholds come from explains why the source and context stay attached.
What an AUC value means
The ranking interpretation is often the cleanest: AUC is the probability that a randomly chosen positive case receives a higher score than a randomly chosen negative case, with tie handling defined by the method.1 AUC = .82 therefore describes ranking performance, not “82% accurate” predictions.
Report a confidence interval or other uncertainty estimate. Two AUC point estimates can differ numerically while remaining statistically and practically indistinguishable on the validation sample.
Class imbalance and case mix
ROC curves are insensitive to prevalence in their algebra, but the usefulness of a model is not. In heavily imbalanced settings, apparently strong ROC performance can coexist with poor positive predictive value at relevant thresholds. Precision–recall summaries can reveal behavior hidden by ROC curves in those settings.4
That observation does not expand this page into the excluded general precision vs recall curriculum. The practical instruction is narrower: inspect prevalence and threshold-specific confusion counts before calling AUC “good.” F1 score interpretation covers another task-dependent classification summary.
Discrimination is not calibration
A model can rank cases correctly while systematically predicting probabilities that are too high or too low. AUC measures discrimination, not calibration. Inspect a calibration plot and a proper scoring rule where probabilities matter; Brier score and calibration covers that distinction.
Also identify what produced the score. In logistic regression, coefficients and odds ratios describe modeled associations, while AUC evaluates ranking; logistic regression output keeps those artifacts separate.
A defensible reporting sentence
“On the held-out sample, AUC was .82 (95% CI [.78, .86]), indicating good ranking performance and falling in the cited .80–.90 convention. Threshold-specific errors, prevalence, and calibration were evaluated separately.”