On linear combinations of biomarkers to improve diagnostic accuracy

Abstract

We consider combining multiple biomarkers to improve diagnostic accuracy. Su and Liu derived the linear combinations that maximize the area under the receiver operating characteristic (ROC) curves. These linear combinations, however, may have unsatisfactory low sensitivity over a certain range of desired specificity. In this paper, we consider maximizing sensitivity over a range of specificity. We first present a simpler proof for Su and Liu's main theorem and further investigate some other optimal properties of their linear combinations. We then derive alternative linear combinations that have higher sensitivity over a range of high (or low) specificity. The methods are illustrated using data from a study evaluating biomarkers for coronary heart disease.

Diffusion-weighted magnetic resonance imaging for the assessment of fibrosis in chronic hepatitis C

“…15 Optimal linear combinations of biomarkers were investigated for improving the diagnosis. 16 Optimal cutoff values for DWMRI were chosen to maximize the sum of the sensitivity and specificity, and the positive and negative predictive values were computed for these cutoff values.…”

Biomarkers for the Diagnosis, Prognosis, and Evaluation of Treatment Efficacy for Traumatic Brain Injury

“…Modifications to this liner combination approach (e.g., in which sensitivity is maximized instead of the area over a range of specificity) have also been described. 150 Another method that could be used to overcome this problem is by applying boosting algorithms. 151 Boosting occurs in stages, during each of which a weak predictor (e.g., a biomarker that has a low specificity) is trained with the data.…”

Plus disease in retinopathy of prematurity: Pilot study of computer-based and expert diagnosis

“…Logistic regression was used to model 26 possible linear combinations of parameters, taken two or more at a time. 19 For each individual parameter and linear combination, sensitivity and specificity of the computer-based system was plotted as a function of threshold used to separate "plus" from "not plus." ROC curves were plotted for individual parameters and linear combinations.…”