Development of a diagnostic test based on multiple continuous biomarkers with an imperfect reference test

Area of Science:

• Biostatistics
• Biomarker Discovery
• Diagnostic Accuracy

Background:

• Reference tests for biomarker accuracy can be imperfect, leading to biased diagnostic accuracy estimates.
• Accurate estimation of diagnostic accuracy is vital for selecting effective biomarker combinations.

Purpose of the Study:

• To propose a Bayesian latent-class mixture model for selecting biomarker combinations that maximize the area under the ROC curve (AUC).
• To develop a method for prior specification that controls information on AUC, accounting for imperfect reference tests.

Main Methods:

• A Bayesian latent-class mixture model was developed and evaluated using simulation studies.
• The model was applied to real-world data from Alzheimer's disease research.
• Prior specifications for mixture component parameters were investigated, including informative and flat priors.

Main Results:

• Simulation studies demonstrated satisfactory performance and accurate parameter estimates for the proposed model.
• The Bayesian model yielded substantially higher AUC estimates compared to traditional logistic regression models that ignore reference test imperfections.
• Model performance was evaluated across various sample sizes and biomarker correlations.

Conclusions:

• The proposed Bayesian model effectively addresses the challenge of imperfect reference tests in biomarker selection.
• This approach provides more reliable and higher diagnostic accuracy estimates than conventional methods.
• The model is valuable for biomarker research, particularly in complex diseases like Alzheimer's.