Development and pilot temporal validation of the PREST score for predicting non-response to intravenous glucocorticoids in active moderate-to-severe thyroid eye disease: a retrospective cohort study

Abstract Background Intravenous glucocorticoids are the first-line treatment for active moderate-to-severe thyroid eye disease, but 20% to 60% of patients respond inadequately, and lower response rates have been reported in Asian cohorts. We aimed to develop and pilot-validate a simple score based on routine baseline measures to predict non-response. Methods In this retrospective single-center cohort study, adults with active (clinical activity score ≥ 3) moderate-to-severe thyroid eye disease who completed 12 weeks of intravenous methylprednisolone (cumulative dose 4.5 g) were divided into a development cohort (2016–2020; n = 89) and a temporal validation cohort (2021–2024; n = 24). Non-response was defined as early treatment escalation for predefined clinical reasons or improvement in fewer than two of five response components. Predictors were selected with logistic regression and converted to integer points. Discrimination (area under the receiver operating characteristic curve, AUC), calibration (Brier score), and decision curve analysis were assessed, and the score was compared with five machine-learning classifiers. Results Non-response occurred in 19 of 89 patients (21.3%) in the development cohort and 6 of 24 (25.0%) in the validation cohort. The PREST score assigned 3 points for thyrotropin receptor antibody ≥ 4.59 IU/L, 2 points for clinical activity score ≥ 4, 2 points for proptosis ≥ 19 mm, and 1 point for thyroid-stimulating immunoglobulin ≥ 406%. The AUC was 0.809 (95% confidence interval 0.696–0.909) in development and 0.870 (0.721–0.974) in validation. At a cutoff of 6 points, sensitivity and specificity were 73.7% and 74.3% in development and 66.7% and 77.8% in validation. The Brier score was 0.159, and decision curve analysis showed net benefit at threshold probabilities of about 10% to 20%. In development, the response rate fell from 96.3% for scores of 0–2 to 41.2% for scores ≥ 7 (P for trend < 0.001). Machine-learning models discriminated less well than PREST in validation. Conclusions A four-item score based on routinely available serologic and ophthalmic measures stratified the risk of non-response to intravenous glucocorticoids in this single-center pilot study. Validation in larger multicenter cohorts is needed before clinical use.