Development and external validation of a nomogram for identifying patients with thyroid eye disease who screen positive for depressive symptoms: a two-center study.

Abstract

Objective: To identify factors independently associated with a positive depressive symptom screen in patients with thyroid eye disease (TED) and to develop and externally validate a nomogram-based model for individualized identification in a second-center cohort as a clinically feasible screening tool for prioritizing depressive symptom screening and informing ongoing follow-up and targeted multidisciplinary management.

Methods: In this two-center observational study, a development cohort (n=218; October 2025-January 2026) and an external validation cohort (n=88; February-March 2026) were enrolled from two tertiary hospitals in Jiangsu Province (total 306 patients). A positive depressive symptom screen was defined as a PHQ-9 total score ≥5. Independent associated factors were selected by univariate and multivariate logistic regression. Five models were developed using the same development data and preprocessing pipeline: logistic regression, LASSO, support vector machine, random forest, and XGBoost. The entire modeling process was internally validated and corrected for optimism using 500 bootstrap resamples, and hyperparameters were optimized using stratified five-fold cross-validation. Discrimination (AUC), decision curve analysis (DCA), and external validation were assessed, and the final locked model was applied to the external cohort without updating.

Results: Overall, 116/306 patients (37.9%) screened positive for depressive symptoms. Age, average monthly household income, stressful life events in the past 6 months, EUGOGO severity grade, standardized GO-QoL-AP score, and TRAb level were independent factors associated with a positive depressive symptom screen. The optimism-corrected AUCs of the five models ranged from 0.736 to 0.790, with no statistically significant paired differences between the corrected AUC of logistic regression and those of the other models. Logistic regression was selected and locked. Its apparent development-cohort AUC was 0.815 and its 500-bootstrap-corrected AUC was 0.776 (95% CI, 0.724-0.840; Brier score, 0.186). In external validation, the AUC was 0.742 (95% CI, 0.631-0.843), the Brier score was 0.195, the calibration slope was 0.981, and the calibration intercept was 0.007. At the locked threshold of 0.312, sensitivity was 0.771 and specificity was 0.547. DCA showed some clinical net benefit across threshold probabilities of 0.10-0.50.

Conclusions: We developed a nomogram for identifying patients with TED who screen positive for depressive symptoms based on age, mean monthly household income, stressful life events during the preceding 6 months, EUGOGO severity grade, standardized GO-QoL-AP score, and TRAb level. In second-center external validation, the model showed moderate discrimination, with calibration slope and intercept estimates close to their ideal values. It may serve as an auxiliary tool for prioritizing depressive symptom screening among patients with TED aged 18-59 years.