Abstract
<title>Abstract</title> <p>Background The clinical evaluation is still the basis for diagnosing thyroid eye disease (TED). Disease activity in most cases is judged by means of the Clinical Activity Score (CAS). Because of inter-rater variability in the score, serum-based biomarkers may provide a complementary source of information for diagnosis and for tracking disease activity. Methods We used a group of 77 individuals, including 54 patients with TED and 23 healthy control (HC) subjects. Serum levels of 25 factors were measured using Meso Scale Discovery (MSD) and enzyme-linked immunosorbent assay (ELISA) platforms. Group comparisons were performed with Mann–Whitney U tests followed by Benjamini–Hochberg correction. Relations among the analytes were investigated by Spearman rank correlation analysis, and diagnostic models were built by LASSO-regularized logistic regression. Model performance was assessed using 5-fold cross-validation (CV), with confidence intervals obtained by bootstrap resampling. Results Of the 25 factors considered, macrophage inflammatory protein-3α (MIP-3α; q = 0.012) and CD44 (q = 0.048) were present at lower levels in patients with TED than in HC. Fourteen factor pairs had an absolute correlation coefficient (|r|) of 0.70 or more. The collinear cluster contained interleukin-21 (IL-21), IL-22, IL-23, IL-27, IL-31, and transforming growth factor-β1. The 4-factor signature, comprising MIP-3α, CD44, IL-31, and IL-8, produced a training AUC of 0.845 (95% CI 0.737–0.934) and a CV AUC of 0.794 (95% CI 0.676–0.892). At the Youden-optimized cutoff of 0.461, sensitivity was 100%, specificity was 60.9%, and overall accuracy was 88.3%. The full 25-factor LASSO model generated a training AUC of 0.816 and a lower CV AUC of 0.622, a difference consistent with overfitting. The probability score from the 4-factor signature was not significantly associated with the Clinical Activity Score (CAS) (r = 0.203, P = 0.140). A separate axis defined by IL-8, tumor necrosis factor-α, CD44, apolipoprotein F, and CD109 was positively correlated with CAS (r = 0.455, P < 0.001) and discriminated active from inactive TED with an AUC of 0.775 (95% CI 0.634–0.884). Conclusions In this patient group, two different serum profiles were identified: a 4-factor diagnostic fingerprint related to TED diagnosis and a 5-factor activity axis related to disease activity assessment. Both biomarker sets remain candidates, and their clinical use needs to be determined in independent patient populations.</p>