Ordinal regression methodology for ROC curves derived from correlated data

Area of Science:

• Medical Statistics
• Diagnostic Imaging Analysis
• Biomedical Data Science

Background:

• Correlated Receiver Operating Characteristic (ROC) data presents analytical challenges in diagnostic accuracy studies.
• Existing methods may not adequately account for multiple interpretations or modalities for the same patient.
• Incorporating patient and reader characteristics is crucial for comprehensive diagnostic performance evaluation.

Purpose of the Study:

• To develop and present a robust statistical approach for analyzing correlated ROC data.
• To enable the incorporation of patient and reader covariates into diagnostic accuracy analyses.
• To demonstrate the utility of the proposed method using real-world diagnostic oncology data.

Main Methods:

• Utilized ordinal regression models combined with generalized estimating equations (GEE).
• Applied the approach to analyze degree-of-suspicion data from multiple interpretations and modalities.
• Integrated patient and reader characteristics directly into the regression models without stratification.

Main Results:

• The proposed method effectively analyzes correlated ROC data from complex diagnostic scenarios.
• Demonstrated successful incorporation of patient and reader covariates, enhancing analytical depth.
• Successfully applied to two diagnostic oncology studies, validating its practical utility.

Conclusions:

• The presented approach offers a powerful tool for analyzing correlated ROC data in medical diagnostics.
• This method enhances the ability to assess diagnostic accuracy while accounting for complex data structures and covariates.
• The findings have significant implications for improving the statistical rigor in diagnostic performance research.