Risk Assessment of Bone Metastasis for Cervical Cancer Patients by Multiple Models: A Large Population Based Real-World Study

Background: Population-based data on the risk assessment of newly diagnosed cervical cancer patients' bone metastasis (CCBM) are lacking. This study aimed to develop various predictive models to assess the risk of bone metastasis via machine learning algorithms. Materials and Methods: We retrospectively reviewed the CCBM patients from the Surveillance, Epidemiology, and End Results (SEER) database of the National Cancer Institute to risk factors of the presence of bone metastasis. Clinical usefulness was assessed by Akaike information criteria (AIC) and multiple machine learning algorithms based predictive models. Concordance index (C-index) and receiver operating characteristic (ROC) curve were used to define the predictive and discriminatory capacity of predictive models. Results: A total of 16 candidate variables were included to develop predictive models for bone metastasis by machine learning. The areas under the ROC curve (AUCs) of the random forest model (RF), generalized linear model (GL), support vector machine (SVM), eXtreme Gradient Boosting (XGBoost), artificial neutral network (ANN), decision tree (DT), and naive bayesian model (NBM) ranged from 0.85 to 0.93. The RF model with 10 variables was developed as the optimal predictive model. The weight of variables indicated the top seven factors were organ-site metastasis (liver, brain, and lung), TNM stage and age. Conclusions: Multiple machine learning based predictive models were developed to identify risk of bone metastasis in cervical cancer patients. By incorporating clinical characteristics and other candidate variables showed robust risk stratification for CCBM patients, and the RF predictive model performed best among these predictive models.

Automated summary

Population-based data on newly diagnosed cervical cancer patients' bone metastasis risk are lacking. This study developed various predictive models using machine learning algorithms to assess the risk of bone metastasis, with the random forest model showing the best performance. Important factors influencing bone metastasis included organ-site metastasis, TNM stage, and age.