Compare diagnostic tests using transformation-invariant smoothed ROC curves()

Area of Science:

• Biostatistics
• Medical Diagnostics
• Machine Learning

Background:

• Receiver operating characteristic (ROC) curves are crucial for evaluating diagnostic biomarkers.
• Existing ROC estimation methods often fail to guarantee essential properties like monotonicity and smoothness.
• Previous methods, such as Du and Tang (2009), do not account for correlations between empirical ROC estimates, complicating asymptotic analysis.

Purpose of the Study:

• To propose a novel penalized weighted least square estimation method for ROC curves.
• To ensure the estimated ROC curves possess desirable properties like monotonicity and smoothness.
• To develop a method that accounts for the inherent correlations among empirical ROC estimates.

Main Methods:

• A penalized weighted least square approach is utilized, incorporating the covariance matrix of empirical ROC estimates as weights.
• The consistency of the proposed estimator is theoretically established.
• A resampling technique is employed for extending the method to compare multiple diagnostic tests.

Main Results:

• The proposed method yields ROC curve estimators that are monotone, smooth, and consistent.
• Simulations demonstrate superior performance compared to existing methods, particularly for steep ROC curves.
• The method was successfully applied to a cancer diagnostic study comparing new biomarkers against a traditional one.

Conclusions:

• The penalized weighted least square method provides a robust and accurate approach for ROC curve estimation in diagnostic studies.
• This method addresses limitations of prior techniques by incorporating covariance information.
• The findings have significant implications for biomarker evaluation and diagnostic test comparisons.