Prediction of 5-year survival in soft tissue leiomyosarcoma using a machine learning model algorithm

Background and Objectives: Leiomyosarcoma (LMS) is associated with one of the poorest overall survivals among soft tissue sarcomas. We sought to develop and externally validate a model for 5-year survival prediction in patients with appendicular or truncal LMS using machine learning algorithms. Methods: The Surveillance, Epidemiology, and End Results (SEER) database was used for development and internal validation of the models; external validation was assessed using our institutional database. Five machine learning algorithms were developed and then tested on our institutional database. Area under the receiver operating characteristic curve (AUC) and Brier score were used to assess model performance. Results: A total of 2209 patients from the SEER database and 81 patients from our tertiary institution were included. All models had excellent calibration with AUC 0.84−0.85 and Brier score 0.15−0.16. After assessing the performance indicators according to the TRIPOD model, we found that the Elastic-Net Penalized Logistic Regression outperformed other models. The AUCs of the institutional data were 0.83 (imputed) and 0.85 (complete-case analysis) with a Brier score of 0.16. Conclusion: Our study successfully developed five machine learning algorithms to assess 5-year survival in patients with LMS. The Elastic-Net Penalized Logistic Regression retained performance upon external validation with an AUC of 0.85 and Brier score of 0.15.