Diagnostic Threshold Identification for Equine Laminitis Using Smoothed Receiver Operating Characteristic Analysis.

Abstract

Radiographic measurement parameters play a key role in diagnosing acute and subacute laminitis in horses, with diagnostic thresholds typically derived from empirical receiver operating characteristic (ROC) analysis. However, these methods often produce unstable results, particularly in small or noisy datasets. This study applies nonuniform rational B-spline (NURBS)-based ROC smoothing to radiographic data from laminitic and healthy horses to evaluate its effectiveness in refining diagnostic thresholds. Thresholds are determined using Youden's index based on NURBS-smoothed ROC curves and compared against the empirical thresholds previously proposed in the literature, which are obtained via the index of union criterion. Diagnostic performance is assessed by area under the curve (AUC), sensitivity, specificity, and threshold agreement, with bootstrap resampling used to estimate confidence intervals. Across multiple radiographic variables, smoothed ROC analysis yields thresholds that match or exceed conventional methods in classification accuracy. NURBS-derived thresholds reduce misclassification rates and show stronger alignment with clinical diagnosis. The use of smoothing and resampling techniques strengthens the statistical reliability of the findings. NURBS smoothing emerges as a reliable complementary tool for improving consistency in radiographic threshold selection for laminitis detection.