Abstract
BACKGROUND: The aim of this study is to develop a machine learning-based model to predict postoperative overall survival (OS) in patients with soft tissue sarcoma (STS) that demonstrates superior comprehensive performance. METHODS: This analysis leveraged data from the SEER database spanning 2010-2020, alongside a STS cohort from the National Cancer Center. Machine learning methods were applied for predictor selection by wrapper methods and the development of the predictive model. The optimal model was determined using the concordance index (C-index), time-dependent calibration curves, time dependent receiver operating characteristic (ROC) curves, and decision curve analysis (DCA). RESULTS: Six machine learning learners identified six feature subsets. Subsequently, six feature subsets and six machine learning learners were combined, resulting in the development of 36 prognostic models. The CAM model, exhibiting the highest prediction performance, was selected. The CAM model achieved a C-index of 0.849 (95% CI 0.837-0.859) in the training cohort and 0.837 (95% CI 0.809-0.871) in the validation cohort. Furthermore, time-dependent calibration curves, time-dependent ROC curves, and DCA indicate that the PAM demonstrates excellent calibration, predictive accuracy, and clinical net benefit. A publicly accessible web tool was developed for the CAM. Notably, CAM's performance exceeds that of all existing STS prognostic nomograms and prediction models. CONCLUSIONS: The CAM has the potential to identify postoperative OS in STS patients. This can assist clinicians in assessing the severity of the disease, facilitating patient follow-up, and aiding in the formulation of adjuvant treatment strategies.