mROC: a computer program for combining tumour markers in predicting disease states.

Area of Science:

• Biostatistics
• Medical Diagnostics
• Statistical Modeling

Background:

• Traditional Receiver Operating Characteristic (ROC) curve analysis faces challenges with multiple diagnostic tests due to multiplicity and inter-test relationships.
• Univariate analyses may incorrectly discard valuable individual markers when evaluating multiple diagnostic tests.
• Existing methods struggle to optimally integrate information from several related diagnostic markers.

Purpose of the Study:

• To present a novel program utilizing generalized ROC criteria to address the limitations of traditional ROC curves with multiple diagnostic tests.
• To propose a method for finding the best linear combination of diagnostic tests that maximizes the area under the ROC curve.
• To demonstrate the utility of data transformation and marker combination for enhancing diagnostic discriminative power.

Main Methods:

• Application of generalized ROC criteria and confidence intervals derived from the non-central F distribution.
• Assumption of multivariate normal distribution for quantified marker values, allowing for unequal variances between populations.
• Inclusion of Box-Cox variable transformations, QQ-plots, and interactive graphics for sensitivity/specificity analysis.

Main Results:

• The proposed method identifies the optimal linear combination of diagnostic tests, maximizing the area under the ROC curve.
• Data transformation techniques, such as Box-Cox, can significantly enhance the discriminative power of diagnostic markers.
• Combining multiple markers through linear combinations demonstrates superior performance compared to individual marker analysis.

Conclusions:

• Generalized ROC criteria offer a robust solution for evaluating multiple diagnostic tests, overcoming limitations of traditional methods.
• Linear combinations of diagnostic markers, optimized via generalized ROC, can substantially improve diagnostic accuracy.
• Relying solely on univariate analyses for marker selection may be suboptimal and lead to the exclusion of informative markers.