Objective A prediction model for distant metastasis in thyroid cancer was developed utilizing data from the Surveillance, Epidemiology, and End Results (SEER) database of the U.S. National Cancer Institute to serve as a quantitative tool to assist clinicians in identifying suitable patients for ¹⁸F-fluorodeoxyglucose (FDG) PET/CT precisely.

Methods This investigation was conducted as a retrospective case-control study. Data were downloaded from 90 934 patients diagnosed with thyroid cancer (comprising 69 716 females and 21 218 males, aged 1–84 years, with a median age of 50 years) from the SEER database, covering the period from 2000 to 2019. Employing stratified random sampling, the patient data derived from SEER database were divided into a training set and a validation set at a ratio of 3∶1. Two prediction models for distant metastasis in thyroid cancer, namely, the Logistic regression model and the stacking ensemble model, were developed using the Logistic regression algorithm and stacking ensemble algorithm, respectively, in the training set. Internal validation was performed using the SEER-based training and validation sets, with primary evaluation metrics including the area under receiver operating characteristic curve (AUROC), the area under precision-recall curve (AUPRC), calibration curve, and Brier score. For external testing, data from 235 patients with thyroid cancer obtained from The Cancer Genome Atlas (TCGA) database, along with data from 205 patients with thyroid cancer admitted to the Department of Thyroid Surgery at Xuzhou Central Hospital between 2022 and 2024 (i.e., local data), were utilized. The DeLong test was employed to compare the differences in AUROC between the Logistic regression model and the stacking ensemble model. Decision curve analysis (DCA) was conducted to determine the candidate decision threshold range for the clinical application of the models. The permutation feature algorithm within the DALEX package of R4.3.2 was applied for the global interpretation of models and to evaluate the contribution of each variable to the prediction outcomes. Furthermore, a web application (APP) was developed using the Shiny package in R4.3.2.

Results The efficacy evaluation results of the Logistic regression model and the stacking ensemble model demonstrated that, in the training set, the AUROCs of the two models were 0.894 and 0.907, respectively (Z=0.163, P=0.103). The AUPRCs were 0.153 and 0.190, and the Brier scores were 0.011 and 0.013. In the validation set, the AUROCs of the two models were 0.881 and 0.883 (Z=0.147, P=0.251), with AUPRCs of 0.155 and 0.153 and Brier scores of 0.010 for both models. External testing utilizing TCGA data yielded AUROCs of 0.838 and 0.833 (Z=0.749, P=0.520), AUPRCs of 0.103 and 0.113, and Brier scores of 0.019 and 0.020. External testing with local data produced AUROCs of 0.866 and 0.907 (Z=2.513, P=0.012), AUPRCs of 0.372 and 0.356, and Brier scores of 0.026 and 0.029. DCA result indicated that the models' candidate decision threshold ranged from 0.001 to 0.240. The permutation feature algorithm analysis revealed that the American Joint Committee on Cancer T stage was the variable with the most significant influence on prediction results in both models. This research successfully developed a multifunctional, web-based APP.

Conclusions A prediction model for distant metastasis in thyroid cancer, designed to identify suitable patients for ¹⁸F-FDG PET/CT, was developed utilizing the SEER database. Internal validation and external testing verified that the model's efficacy adhered to clinical standards. The model is anticipated to improve screening efficiency for appropriate patients undergoing ¹⁸F-FDG PET/CT and facilitate personalized diagnosis and treatment of patients with thyroid cancer.