ROC Curve

The ROC curve displays the sensitivity (true positive rate) versus 1-specificity (false positive rate) of a binary classifier at all possible classification thresholds. The area under the curve (AUC) quantifies overall discrimination. A ROC curve without AUC and its 95% CI is incomplete —the AUC alone without CI does not allow comparison between classifiers or assessment of clinical utility.

STEP 1 —When to Use

Use a ROC curve when: you have a binary outcome (disease / no disease, event / no event) and a continuous predictor (biomarker, model score, probability); you want to evaluate the discriminative ability of the predictor regardless of threshold; you want to compare two classifiers (DeLong test for AUC comparison).

AUC interpretation: AUC = 0.5 is random chance (diagonal line); AUC = 1.0 is perfect discrimination. Common thresholds for clinical use: >= 0.70 acceptable, >= 0.80 good, >= 0.90 excellent. But AUC should always be reported with its 95% CI —an AUC of 0.75 with CI [0.61, 0.89] is substantially more uncertain than one with CI [0.73, 0.77].

[Image placeholder: ROC curve with the diagonal reference line (AUC=0.5). The ROC curve shown in blue, AUC=0.84 annotated in the upper left. Optimal threshold point (Youden index) marked as a filled circle on the curve. Sensitivity and specificity at that threshold labeled as text annotations. 95% CI for AUC shown as a shaded region under the curve.]

ROC curve with AUC annotation, optimal threshold point (Youden index), sensitivity/specificity at threshold, and 95% CI for AUC.

STEP 2 —Journal Requirement

AUC with 95% CI (DeLong method) must be reported.
n (cases and controls) must be stated.
If an optimal threshold is claimed: the basis for threshold selection must be stated (Youden index, sensitivity constraint, specificity constraint).
Sensitivity and specificity at the reported threshold with 95% CI must be stated.
For AUC comparison between two classifiers: DeLong test p-value required.

STEP 3 —Licklider Implementation

Input

Binary outcome column (0/1)
Continuous predictor/score column
Optional: second predictor for comparison curve
Threshold selection: Youden index (default), sensitivity constraint (e.g., >=90% sensitivity), or user-specified

Output

ROC curve (1-specificity on x-axis, sensitivity on y-axis)
Diagonal reference line (AUC = 0.5)
AUC with 95% CI (DeLong) annotated
Optimal threshold marked as a point on the curve
Sensitivity and specificity at optimal threshold annotated
95% CI for AUC shown as shaded envelope under the curve (optional)
For comparison: second curve in distinct color, DeLong test p-value annotated

[Image placeholder: Licklider ROC curve for 2 classifiers (blue: Biomarker A, orange: Biomarker B). Reference diagonal dashed. AUC annotation: "Biomarker A: AUC = 0.84 (95% CI [0.74, 0.94]); Biomarker B: AUC = 0.71 (95% CI [0.59, 0.83]); DeLong p = 0.04." Optimal threshold marked on each curve. n=80 cases, 80 controls labeled.]

Licklider ROC: two classifiers compared, AUC with CI for each, DeLong comparison p-value, optimal thresholds marked.

STEP 4 —Draft Output (Draft / Needs review)

ROC curve analysis for [Biomarker] in distinguishing [cases] from [controls]
(n = 80 cases, 80 controls). AUC = 0.84 (95% CI [0.74, 0.94], DeLong method).
The optimal classification threshold (Youden index: sensitivity + specificity ~1
maximized) was [threshold value], yielding sensitivity = 78% (95% CI [68, 86])
and specificity = 81% (95% CI [71, 89]). The diagonal line represents the
performance of a random classifier (AUC = 0.50).