Bayesian multidimensional nominal response model for observer study of radiologists.

Area of Science:

• Statistical modeling
• Multiclass classification analysis
• Bayesian inference

Background:

• Multiclass classification problems require robust statistical models.
• Existing nominal response models may face convergence challenges.
• Accurate estimation of latent parameters is crucial for model interpretation.

Purpose of the Study:

• To propose a Bayesian multidimensional nominal response model (MD-NRM) for multiclass classification.
• To extend the conventional nominal response model for stable Bayesian convergence.
• To apply MD-NRM to a real-world chest X-ray classification task.

Main Methods:

• Developed and applied the Bayesian multidimensional nominal response model (MD-NRM).
• Utilized multidimensional ability parameters and a softmax function assumption.
• Estimated latent parameters (radiologist ability, image difficulty) using Stan (version 2.21.0).

Main Results:

• The Bayesian MD-NRM demonstrated stable convergence, with Rhat values < 1.10 for all parameters.
• Successfully estimated latent ability and difficulty parameters from 900 nominal responses.
• The model proved effective in a 3-class chest X-ray diagnosis classification problem.

Conclusions:

• Latent parameters of the Bayesian MD-NRM can be reliably estimated using Stan.
• The developed MD-NRM provides a statistically sound approach for multiclass classification.
• Open-source code for the MD-NRM implementation is available, facilitating further research.