Smooth non-parametric receiver operating characteristic (ROC) curves for continuous diagnostic tests

Area of Science:

• Biostatistics
• Medical Informatics
• Diagnostic Test Evaluation

Background:

• Receiver operating characteristic (ROC) curves are crucial for evaluating diagnostic test accuracy with continuous results.
• Traditional non-parametric ROC curves can be jagged, while parametric methods risk model misspecification.
• Existing smooth alternatives like LABROC4 lack full flexibility.

Purpose of the Study:

• To propose a smooth, non-parametric ROC curve method using kernel density estimates.
• To develop methods for constructing confidence intervals and rectangles for ROC curves and areas.
• To compare the proposed method with existing techniques using real-world datasets.

Main Methods:

• Kernel density estimation applied to continuous test results to derive smooth ROC curves.
• Calculation of pointwise standard errors for true positive rates (TPR) and false positive rates (FPR) for confidence intervals.
• Adaptation of existing methods for calculating area under the ROC curve (AUC) and its standard error.
• Development of a FORTRAN algorithm named 'ROC-&-ROL'.

Main Results:

• The proposed method generates smooth, non-parametric ROC curves.
• Pointwise confidence intervals and confidence rectangles provide robust uncertainty quantification.
• The new method demonstrates superior performance on a dataset where parametric models fail.
• Comparison with existing methods highlights the flexibility and accuracy of the proposed approach.

Conclusions:

• The kernel density-based smooth non-parametric ROC curve method offers a flexible and reliable alternative for diagnostic accuracy assessment.
• This approach mitigates issues associated with jagged curves and parametric model assumptions.
• The developed 'ROC-&-ROL' algorithm facilitates the practical application of these advanced statistical techniques.