Retrospective five-variable model predicts Graves’ relapse after drug withdrawal

A nomogram built from five routinely available baseline measures identified adults with Graves’ disease who were most likely to relapse within one year of stopping antithyroid drugs. In a retrospective cohort of 654 patients treated at Dazhou Central Hospital in China, the model combined age, sex, goiter grade, free thyroxine and thyrotropin receptor antibody levels to generate an individualized probability of relapse. It achieved an area under the receiver operating characteristic curve of 0.84 in a held-out testing set, indicating good discrimination between patients who relapsed and those who remained in remission. The results provide a promising risk-assessment tool, although the authors emphasize that independent validation in larger and more diverse populations is required before routine clinical use.

The single-center observational study included adults newly diagnosed with Graves’ disease who received methimazole or propylthiouracil as initial treatment, achieved stable remission on a minimal maintenance dose after at least 12 months of therapy, and then stopped the medication. Remission required restoration and stable maintenance of normal thyroid-stimulating hormone, free triiodothyronine and free thyroxine within their reference ranges. Relapse was defined as biochemical overt thyrotoxicosis, meaning suppressed thyroid-stimulating hormone together with elevated free thyroxine and/or free triiodothyronine, or as recurrence of typical hyperthyroid symptoms leading to renewed or increased antithyroid treatment, radioactive iodine or thyroidectomy. A single qualifying biochemical measurement was sufficient under the study definition, without a required repeat test or minimum interval between measurements.

Among the 654 participants, 397 experienced relapse within one year and 257 maintained remission, producing an overall relapse proportion of 60.7%. Patients who relapsed were generally younger and more likely to be male, have a larger goiter and have higher concentrations of thyroid hormones and thyroid antibodies. Their median age was 32 years, compared with 48 years among those who did not relapse. Grade 2–3 goiter was present in 56.2% of the relapse group but only 22.8% of the non-relapse group. Median baseline free thyroxine was 78.30 pmol/L in patients who later relapsed versus 51.50 pmol/L in those who remained in remission, while median thyrotropin receptor antibody levels were 21.02 and 8.40 IU/L, respectively. These descriptive differences were measured in the retrospective cohort and do not by themselves establish causal effects.

Researchers used least absolute shrinkage and selection operator regression, or LASSO, to screen a broad set of demographic, clinical and laboratory variables in a training partition containing 70% of the cohort. Eight variables initially received nonzero coefficients: sex, age, goiter grade, free thyroxine, thyroid-stimulating hormone, initial drug type, time from symptom onset to diagnosis and thyrotropin receptor antibody. A multivariable logistic regression, fitted across five multiply imputed datasets, then reduced these to five independent predictors. Younger age and male sex were associated with greater relapse odds, as were grade 2–3 goiter, higher free thyroxine and higher thyrotropin receptor antibody concentration. Each additional year of age was associated with an adjusted odds ratio of 0.96, or 4% lower odds, while the adjusted odds ratio was 1.88 for grade 2–3 goiter compared with grade 0–1. The odds ratio for female sex relative to male sex was 0.49, with a 95% confidence interval of 0.27–0.87.

Each 1 pmol/L increase in baseline free thyroxine was associated with an adjusted odds ratio of 1.01, while each 1 IU/L increase in thyrotropin receptor antibody was associated with an adjusted odds ratio of 1.08. Time from symptom onset to diagnosis, thyroid-stimulating hormone and initial drug type did not remain independently significant after mutual adjustment. A higher baseline ratio of free triiodothyronine to free thyroxine was associated with relapse in a univariable comparison, but it was not retained by LASSO and adding it to the final model did not improve discrimination: the change in area under the curve was −0.014, with a 95% confidence interval of −0.043 to 0.002. The study also used multiple imputation for retained predictors in the training set and excluded variables with more than 30% missing data from model construction. Ten of 198 patients in the testing partition had at least one missing final predictor and could not receive a predicted probability for the receiver operating characteristic analysis.

To visualize the model, the investigators translated the five regression coefficients into a nomogram in which points assigned to age, sex, goiter grade, free thyroxine and thyrotropin receptor antibody can be summed and mapped to a relapse probability. They also created an interactive online calculator. In the 30% testing partition, patients divided into quartiles by predicted risk showed a strong gradient in observed outcomes. Relapse occurred in 23.4% of the lowest-risk quartile, 42.6% in the second, 76.6% in the third and 97.9% in the highest, with a statistically significant trend across the four groups. At a predicted probability of at least 20%, the model identified 96.5% of patients who relapsed, with a specificity of 14.7%, positive predictive value of 63.0% and negative predictive value of 73.3%. At a threshold of at least 40%, the corresponding values were 90.3%, 45.3%, 71.3% and 75.6%. These are associations and predictive measurements within the study cohort, not evidence that changing any one predictor would causally change an individual patient’s outcome.

Performance was assessed through discrimination, calibration and decision curve analysis. The model’s area under the curve was 0.82 in the training set, with a 95% confidence interval of 0.78–0.86, and 0.84 in the testing set, with a 95% confidence interval of 0.79–0.90. Calibration plots showed agreement between predicted probabilities and observed relapse proportions, although minor fluctuations appeared in the lower-to-moderate probability range of the testing set, where fewer patients contributed to individual calibration bins. Bootstrap internal validation with 1,000 resamples produced an optimism-corrected concordance index of 0.82. The testing-set calibration intercept was −0.17, with a 95% confidence interval of −0.53 to 0.18, and the calibration slope was 1.09, with a 95% confidence interval of 0.77–1.42. Decision curve analysis indicated positive net benefit over treating everyone as high risk or no one as high risk across a range of relevant threshold probabilities, but this metric does not itself establish that using the model improves patient outcomes.

In a direct comparison using the same 188 complete testing cases, the new model performed similarly to the established Graves Recurrent Events After Therapy, or GREAT, score. The new nomogram had an area under the curve of 0.84, compared with 0.82 for GREAT, and the difference was not statistically significant by the DeLong test, which gave a p value of 0.31. The bootstrap-estimated mean difference was 0.018, with a 95% confidence interval of −0.019 to 0.055. The GREAT score uses fixed categories for age, free thyroxine, thyrotropin-binding inhibitory immunoglobulin and goiter grade, whereas the new approach treats free thyroxine and antibody values continuously and returns a probability rather than a discrete point total. The GREAT classes also separated patients in this cohort: observed relapse rates were 20.0% in class I, 43.3% in class II and 85.7% in class III.

Several limitations constrain how the findings should be interpreted. This was a retrospective, single-center observational study, so unrecorded confounding and selection bias cannot be excluded. Participants were all Chinese Han patients from an iodine-sufficient region, limiting direct transportability to other ethnic, geographic or iodine-status populations. Smoking history, end-of-treatment antibody measurements, detailed ultrasound volumes, comorbid conditions and concomitant medications were unavailable, and the antibody assay measured thyrotropin-binding inhibitory immunoglobulin rather than functional thyroid-stimulating antibody activity. Goiter grading was performed by one outpatient physician rather than independent raters, and inter-observer variability was not assessed. Among the 10 excluded testing patients, seven relapsed, compared with 113 of 188 included complete cases; the difference was not statistically significant, but the excluded group was too small for a robust comparison. The observed relapse rate was also higher than commonly reported estimates, which may reflect the strict remission requirement, the single-center population and the broad outcome definition. The tool therefore provides quantitative risk information for individualized discussion, but it does not establish a treatment policy or prove that nomogram-guided decisions produce better clinical outcomes. Prospective, multicenter external validation is needed to determine whether its performance and calibration persist in diverse settings before it can be recommended for widespread implementation.

Subject of Research: One-year prediction of Graves’ disease relapse after antithyroid drug withdrawal

Article Title: A nomogram based on baseline predictors for Graves’ disease relapse after antithyroid drug withdrawal: a retrospective study

Article References: Dai, R., Li, M., Chen, Y., Wang, J., & Xiong, H. (2026). A nomogram based on baseline predictors for Graves’ disease relapse after antithyroid drug withdrawal: a retrospective study. Frontiers in Endocrinology, 17, Article 1968560. https://doi.org/10.3389/fendo.2026.1968560

Image Credits: AI Generated

DOI: 10.3389/fendo.2026.1968560

Keywords: Graves’ disease, antithyroid drugs, relapse prediction, risk nomogram, thyrotoxicosis, free thyroxine, thyrotropin receptor antibodies, retrospective cohort, model validation, GREAT score, nomogram, based