Analyzing a portion of the ROC curve

Area of Science:

• Medical imaging analysis
• Statistical modeling in diagnostics
• Receiver Operating Characteristic (ROC) curve analysis

Background:

• The area under the ROC curve (AUC) is a standard metric for evaluating diagnostic test performance.
• Comparing ROC curves is crucial, but interest may be limited to specific ranges of false-positive rates.
• Existing methods may not adequately address comparisons focused on partial ROC curve areas.

Purpose of the Study:

• To propose a numerical integration method for calculating and comparing partial areas under ROC curves.
• To derive variance estimates for these partial area comparisons.
• To demonstrate the method's applicability to different data types and sample structures.

Main Methods:

• Numerical integration to compute AUC over a defined range of false-positive rates.
• Derivation of variance estimates for partial AUC comparisons.
• Application to binormal data (continuous or rating scale) from independent or dependent samples.

Main Results:

• The proposed numerical integration method effectively evaluates partial areas under ROC curves.
• Variance estimates were successfully derived, enabling statistical comparisons.
• The method was demonstrated using computed tomographic scan data, comparing scans with and without clinical history.

Conclusions:

• Numerical integration provides a robust approach for comparing specific segments of ROC curves.
• This method enhances the assessment of diagnostic performance when only a relevant range of false-positive rates is of interest.
• The technique is versatile and applicable to various clinical and research scenarios in medical diagnostics.