A novel estimator for the two-way partial AUC.

Area of Science:

• Biostatistics
• Machine Learning
• Medical Diagnostics

Background:

• Two-way partial AUC quantifies classifier performance within restricted sensitivity and specificity ranges.
• Existing tpAUC estimators are computationally intensive for large datasets, hindering bootstrap comparisons.
• Nonparametric trimmed Mann-Whitney U-statistic has high computational complexity (O(n^2)).

Purpose of the Study:

• Develop a novel, computationally efficient estimator for the two-way partial AUC.
• Enable reliable comparison of diagnostic tests and classifiers on large datasets.
• Reduce computational burden for statistical comparisons relying on bootstrapping.

Main Methods:

• Derived a new estimator leveraging graphical and probabilistic AUC representations.
• Implemented using the pROC R package with a trapezoidal rule approach.
• Computational complexity reduced to O(n).

Main Results:

• Proposed estimator demonstrates reduced bias compared to the original tpAUC method.
• Significantly faster computation times for large sample sizes due to O(n) complexity.
• Empirical evaluation via simulations and real-world datasets confirmed performance.

Conclusions:

• The novel estimator improves two-way partial AUC computation efficiency.
• Facilitates comparative analysis of diagnostic tests and classifiers in large-scale studies.
• Overcomes computational bottlenecks of previous methods for bootstrap-based comparisons.