Application of the skew exponential power distribution to ROC curves

Area of Science:

• Biostatistics
• Medical Diagnostics
• Biomarker Discovery

Background:

• Receiver Operating Characteristic (ROC) models, particularly the bi-Normal ROC model, are standard for evaluating biomarker discriminatory ability in medical research.
• Clinical biomarkers often exhibit non-Normal distributions (e.g., skewed, heavy-tailed), which can bias traditional model-based decision thresholds despite an unbiased Area Under the Curve (AUC).

Purpose of the Study:

• To propose and evaluate a novel ROC model based on the skew exponential power (SEP) distribution to accommodate non-Normal biomarker distributions.
• To correct for bias in decision threshold estimation inherent in the bi-Normal ROC model when applied to non-Normal data.
• To assess the utility of the SEP distribution in determining the appropriateness of the bi-Normal model.

Main Methods:

• Development of an ROC model utilizing the skew exponential power (SEP) distribution, which includes parameters for modeling various non-Normal distributions.
• Comparative analysis of the proposed SEP-based ROC model against the traditional bi-Normal ROC model and a non-parametric approach.
• Evaluation through a simulation study and application to a real-world dataset concerning Klebsiella pneumoniae infections.

Main Results:

• The SEP-based ROC model demonstrated efficiency gains in estimating the Area Under the Curve (AUC) compared to existing methods.
• Improved classification rates were achieved using decision cut-points derived from the SEP ROC model.
• The SEP distribution proved effective in assessing the suitability of the bi-Normal model for specific datasets.

Conclusions:

• The proposed SEP-based ROC model offers a robust alternative for biomarker evaluation, particularly when dealing with non-Normal data distributions.
• This approach provides more accurate decision thresholds and improved classification performance in clinical settings.
• The SEP ROC model is recommended for biomarker analysis in the presence of non-Normal data to enhance diagnostic accuracy.