A machine learning-based model for the prediction of thyroid eye disease with oxidative stress-related biomarkers

Thyroid eye disease (TED), the most common adult orbital disease, can significantly impair patients' quality of life. Currently, effective diagnostic and predictive models for TED remain limited, making early intervention and personalized treatment for patients challenging. Oxidative stress (OS) plays an important role in pathogenesis of TED, and OS related biomarkers may serve as good candidates for TED prediction. Here, we integrated the peripheral blood bulk-RNA sequencing data and clinical features of 152 TED, 61 health control (HC), and 20 patients with simple Graves' disease (GD) to identify potential biomarkers. The intersection of TED-HC and TED-GD differentially expressed genes (DEGs) identified 1220 genes strongly correlated with TED. Enrichment analysis showed upregulation of OS-related biological processes in patients with TED. Integration of DEGs, WGCNA results, and OS-related genes identified six genes as candidate biomarkers. Machine learning algorithms suggested three critical candidate genes (KLF2, SELENON, TXNRD1) with high predictive value and were used to construct an oxidative stress-related predictive gene score (OSRPGS). Receiver operating characteristic curve confirmed the predictive value of OSRPGS with an AUC value of 0.733 (TED vs HC) and 0.705 (TED vs GD). Further patient stratification analysis confirmed that the OSRPGS was associated with tear secretion dysfunction. Furthermore, immune infiltration analysis suggested an upregulation of innate immune responses, especially the monocytes/macrophages subtypes, indicating the initiation of OS-related inflammation. Collectively, our study provides a reliable tool for TED prediction and risk assessment based on OS-related biomarkers. OSRPGS may help with the early recognition and intervention in patients with TED.