Development and validation of a machine learning model and nomogram for predicting lung metastasis in thymoma: A population-based study

The lung is the most frequent site of metastasis in patients with advanced thymoma. As a result, predicting the probability of lung metastasis and survival time has become a key focus in clinical practice. However, accurate prediction models are currently lacking. We collected complete data from 2004 to 2015 in the Surveillance, Epidemiology and End Results (SEER) database. Cox regression analysis was used to evaluate factors associated with the prognosis of thymoma patients. These independent prognostic factors were incorporated into the nomogram prognostic model. We validated the survival model by incorporating prognostic factors using receiver operating characteristic curves, calibration plots, C-index, and decision curve analysis curves. Synthetic minority over-sampling technique statistical methods were used to evaluate the machine learning model most suitable for building a predictive model of lung metastasis. Receiver operating characteristic curves, probability density plots, and confusion matrices were used to evaluate the performance of the prediction model. A total of 1776 thymoma patients were included in this study. Among them, age, sex, World Health Organization classification, Masaoka_Koga_Stage, surgery, chemotherapy, and lung metastasis were independent factors affecting metastasis. The machine learning model built by the gradient boosting machine algorithm performed best through internal data validation. Then, we used the gradient boosting machine algorithm to construct a web predictor of lung metastasis, and completed the construction and verification of the nomogram model. We developed an accessible online prediction tool (https://lung-metastasis---thymoma-ckfrve2aqrqr8tsdwbrqik.streamlit.app/) to evaluate the risk of lung metastasis. In addition, a predictive nomogram model was developed to predict the prognosis of patients with lung metastasis.