Machine Learning–based Identification of Prognostic Factors for Surgical Management in Patients With NOS Sarcoma

Figures

• 

  Fig. 1.. Predictive factors for the development of metastasis. A, The decision tree (CART) lists the most important dichotomic features for the occurrence of metastasis in the cohort. The root node on the top presents the most important dichotomic feature. All patients 34.5 years or younger developed metastasis in this cohort. Training accuracy is 0.76, and testing accuracy is 0.77. B, The SHAP value quantifies the impact of a specific feature value on a model prediction by measuring the expected difference in prediction compared with a baseline value. Age and tumor width were ranked important.

• 

  Fig. 2.. Predictive factors for the development of a relapse. A, The decision tree (CART) lists the most important dichotomic features for the development of a relapse in the cohort. The root node on the top presents the most important dichotomic feature. No patient with a BMI less than or equal to 21. 28 kg/m2 developed a relapse in this cohort. Training accuracy is 0.75, and testing accuracy is 0.77. B, The SHAP value quantifies the impact of a specific feature value on a model prediction by measuring the expected difference in prediction compared with a baseline value. Tumor grading and BMI were ranked important.

• 

  Fig. 3.. Predictive factors for survival. A, The decision tree (CART) lists the most important dichotomic features for the development of a relapse in the cohort. The root node on the top presents the most important dichotomic feature. Tumor diameter, tumor depth, and localization at the proximal extremity (0) were the most predictive factors for overall survival. Training accuracy is 0.77, and testing accuracy is 0.54. B, The SHAP value quantifies the impact of a specific feature value on a model prediction by measuring the expected difference in prediction compared with a baseline value. Tumor width, patient age, and BMI were ranked important.