Development of a Support Vector Machine-Based Tool for Survival Prediction after Stereotactic Body Radiotherapy for Liver Metastases Using Multi-Institutional Data

Abstract

<h3>Purpose/Objective(s)</h3> Stereotactic body radiotherapy (SBRT) has been increasingly used as well-tolerated, effective treatment for metastatic liver lesions. Machine learning can predict outcomes and aid in developing treatment regimens; however, few studies have attempted to predict liver SBRT outcomes and most have been limited to a single institution. Support vector machine (SVM) algorithms can address complex classification by mapping data into multi-dimensional space. This study aims to develop an SVM-based model to predict survival after SBRT for liver metastases using multi-institutional data. <h3>Materials/Methods</h3> The Comprehensive Liver Pooled SBRT Outcomes (CaLiPSO) is an international consortium of four high-volume SBRT academic centers. Patients who received SBRT for liver metastases were identified from the CaLiPSO database; baseline demographic, clinical, and dosimetric data were collected. Outcomes of interest were overall survival at 12, 18, and 24 months (12mo-OS, 18mo-OS, and 24mo-OS). We investigated a SVM algorithm using a Gaussian kernel and 10-fold cross-validation schema. Performance was measured by accuracy defined as the ratio of correct classifications to samples and by computing the area under the receiving operator characteristic curve (AUC of ROC). Training was performed using a programming environment. <h3>Results</h3> A total of 132 patients with metastatic liver lesions treated with SBRT were included in this study, 77 females (58.3%) and 55 males (41.7%) with median age of 65 years (range 27-88). The median 2-Gy Equivalent Dose (EQD2) was 60 Gy. At the time of SBRT, 62 patients (47.0%) had extrahepatic disease and 70 (53.0%) did not. The number of patients alive at 12, 18, and 24 months were 64 (48.5%), 52 (39.4%), and 41 (31.1%), respectively. Patients were randomly divided into training (n=92) and testing (n=40) sets. Predictive factors included in the SVM were age, EQD2, planning target volume (PTV), and percentage of normal liver volume. Training set AUCs were above 0.95; results for the testing set are listed in Table 1. Performance at 18 and 24 months had AUC of 0.74 and 0.68, respectively, and was improved compared to at 12 months. <h3>Conclusion</h3> Using clinical and dosimetric variables, promising results were obtained for accurate prediction of liver SBRT outcomes. To our knowledge, this is the first SVM-based tool for SBRT outcome prediction in liver metastases validated using international, multi-institutional data for over 100 patients. This tool can help improve care and facilitate appropriate treatment escalation; it can be implemented online for the radiation oncology community as it uses commonly collected data and is easy to compute. The tool can be further refined and validated with larger cohorts in prospective trials.