A PAUC-based estimation technique for disease classification and biomarker selection.

Area of Science:

• Biostatistics
• Medical Informatics
• Machine Learning

Background:

• The partial area under the receiver operating characteristic curve (PAUC) is crucial for evaluating biomarker combinations in disease classification.
• PAUC quantifies sensitivity within specific false positive rate ranges, vital for medical screening test development.
• Few methods exist that optimize biomarker combinations using PAUC as an objective function.

Purpose of the Study:

• To introduce a novel boosting method for deriving biomarker combinations based explicitly on the PAUC criterion.
• To address high-dimensional settings where biomarker count exceeds observations.
• To integrate stability selection for robust variable selection and sparse prediction rules.

Main Methods:

• A boosting algorithm was developed with the PAUC as the objective function.
• The method is designed for high-dimensional data (more biomarkers than samples).
• Incorporated stability selection for identifying relevant biomarkers and ensuring prediction rule sparsity.

Main Results:

• The proposed method demonstrated strong performance in both variable selection and prediction accuracy on simulated and real data.
• It outperforms established techniques in disease classification when focusing on a limited range of specificity values.
• The method effectively identifies sparse prediction rules using only relevant biomarkers.

Conclusions:

• The developed boosting method offers an effective approach for biomarker combination selection using the PAUC criterion.
• This method is particularly advantageous for medical screening applications requiring specific performance ranges.
• It provides a robust solution for high-dimensional data, enhancing prediction accuracy and interpretability.