Bayesian nonparametric estimation of ROC surface under verification bias.

Area of Science:

• Biostatistics
• Medical Diagnostics
• Machine Learning

Background:

• Receiver Operating Characteristic (ROC) curves assess diagnostic accuracy.
• ROC surfaces generalize ROC curves for three-category diagnostic tests.
• Verification bias, where not all true classes are verified, complicates accuracy assessment.

Purpose of the Study:

• To develop a method for estimating ROC surfaces under verification bias.
• To incorporate covariate information for a comprehensive understanding of diagnostic accuracy.
• To propose a robust Bayesian nonparametric approach.

Main Methods:

• Utilized Dirichlet process mixture priors to model underlying distributions.
• Developed a robust computing algorithm with a missing at random assumption.
• Adapted the method for cases without verification bias, enabling fast computation via Bayesian bootstrap.

Main Results:

• The proposed Bayesian method generally outperforms existing approaches in simulations.
• Human epididymis protein 4 shows better diagnostic ability than CA125 for epithelial ovarian cancer.
• Serum albumin demonstrates prognostic ability in distinguishing hepatocellular carcinoma stages.

Conclusions:

• The novel Bayesian nonparametric approach effectively estimates ROC surfaces with verification bias.
• The method provides valuable insights into diagnostic accuracy and prognostic ability in clinical settings.
• This approach enhances the evaluation of biomarkers for various cancers.