Machine Learning–Based Cytokine Endotyping of Thyroid Eye Disease

Purpose: Thyroid eye disease (TED) exhibits profound phenotypic heterogeneity that conventional binary thyrotropin receptor antibody fails to capture. We aimed to deconstruct this heterogeneity by establishing a cytokine-derived molecular taxonomy. Methods: We profiled 102 plasma cytokines in 101 TED patients using a Luminex assay. Integrative machine learning identified key cytokine signatures. Unsupervised K-means clustering defined four discrete endotypes. A diagnostic nomogram (named the cytokine-derived cluster [CKC] score) was developed and evaluated in an independent external cohort. Results: Machine learning identified a robust four-cytokine signature (vascular cell adhesion molecule 1, C-C motif chemokine ligand 19, C-X3-C motif chemokine ligand 1, chemokine (C-X-C motif) ligand 13). This signature stratified TED patients into four distinct endotypes: CKC1 (inflammatory; n = 34) showed the highest Clinical Activity Score and gaze-evoked pain (23.5%); CKC2 (proliferative; n = 32) demonstrated prominent structural changes with the highest proptosis and diplopia (81.3%); CKC3 (fibrogenic; n = 20) exhibited chronic fibrotic features; and CKC4 (metabolic; n = 15) presented with metabolic comorbidities (diabetes 53.3%). The CKC score nomogram successfully predicted endotypes assignment, achieving an internal area under the receiver operating characteristic curve of 0.742 and maintaining robust discriminative power in the external validation cohort. Conclusions: Using machine learning, we identified a four-cytokine signature that robustly stratifies TED patients into four immunologically and clinically distinct endotypes. The CKC scoring model, derived from this signature, enables the accurate and reproducible prediction of endotype membership.