Abstract
BACKGROUND: Differentiated thyroid cancer (DTC) progresses slowly, but patients with lung metastasis (LM) have a poor prognosis. The aim of this study was to develop and evaluate the predictive ability of machine learning (ML) models in estimating the risk of LM in patients with DTC and to identify the independent risk factors specific to different age and gender subgroups. METHODS: The demographic and clinicopathological data of patients with DTC were obtained from two databases: firstly, the National Institutes of Health Surveillance, Epidemiology, and End Results (SEER) database [2010-2015], which provides extensive epidemiological and clinical information on cancer patients; secondly, the Zhangzhou Municipal Hospital Affiliated to Fujian Medical University [2014-2017], which focuses more on patients' specific clinicopathological characteristics and treatment outcomes. Common variables from both databases were extracted. The data were then split into training, testing and validation sets. The training set was used to build and train ML models, while the testing and validation set were employed to assess the performance of these models. In terms of model development, we established five different ML models: logistic regression (LR), random forest (RF), decision tree (DT), extreme gradient boosting (XGBoost), and gradient boosting machine (GBM). For model validation, we utilized various evaluation metrics, including accuracy, precision, recall, F1 score, Brier score, area under the receiver operating characteristic (ROC) curve (AUROC), area under the precision-recall (PR) curve (PR-AUC), calibration curve, and decision curve analysis (DCA). The importance of various features was ranked and visualized for the top-performing models. RESULTS: The analysis identified age, gender, tumor size, T stage, N stage, and histologic type as significant independent risk factors for LM. The effects of gender, T stage, and histological type on the risk of LM varied across the different age subgroups. In the female population, tumor size was an independent risk factor for LM, while it was not in the male population. GBM achieved an AUROC of 0.982, a Brier score of 0.047, an accuracy of 0.818, and an F1 score of 0.818 in the validation set, outperforming the other models. CONCLUSIONS: The GBM model emerged as an effective tool for identifying high-risk LM populations in DTC, with the potential to guide clinical practice and facilitate the development of individualized treatment plans. Further research to validate these findings across more diverse patient populations and clinical settings is recommended.