Comparison of Performance by Validating Survival Prediction Models Trained with Machine Learning or Statistical Regression for Patients with Spinal Metastases Receiving Radiotherapy

Abstract

<h3>Purpose/Objective(s)</h3> Well-performing survival prediction models (SPMs) help patients and radiation oncologists to choose treatment aligning with prognosis. This retrospective study aims to compare and validate the performance of Metastases location, Elderly, Tumor primary, Sex, Sickness/comorbidity, and Site of radiotherapy (METSSS) model, New England Spinal Metastasis Score (NESMS), and Skeletal Oncology Research Group machine learning algorithm (SORG-MLA) for spinal metastases (SM). <h3>Materials/Methods</h3> From 2010 to 2018, patients who received radiotherapy (RT) for SM at a single tertiary center were enrolled for medical record review. Survival was defined as the interval between the first course RT for SM and death by any cause. Logistic and Cox proportional hazards regression analyses were used to assess the association between variables and survival. The area under receiver-operating characteristics curve (AUROC), Brier score, and decision curve analysis were used to evaluate the performance of SPMs. <h3>Results</h3> A total of 2,786 patients were included for analysis. The three most common tumor primary sites were lung (n=1,130), breast (n=343), and liver (n=279). The most common RT regimen was conventional palliative dose-fractionation (3 Gy × 10 ± 10%, n=1,408, 50.9%) compared to hypo-fractionation (4–5 Gy × 5–6, n=805, 29.1%). The 90-day and 1-year survival rates after RT were 70.4% and 35.7%, respectively. Higher albumin, hemoglobin, or lymphocyte count, while lower alkaline phosphatase, white blood cell count, neutrophil count, neutrophil-to-lymphocyte ratio, platelet-to-lymphocyte ratio, or international normalized ratio were significantly associated with better survival after RT. Patients with serum albumin level ≥ 3.6 g/dL had an odds ratio of 5.05 (95% CI 4.26–6.03) to survive 1-year and a hazard ratio of 0.40 (95% CI 0.38–0.43) to die after RT for SM. SORG-MLA has the highest discriminatory ability (AUROC 90-day 0.78; 1-year 0.76) compared to NESMS and METSSS model (Table, DeLong's test, p<0.001) and the lowest Brier score (90-day 0.16; 1-year 0.18) compared to the null model (90-day 0.21; 1-year 0.23). The decision curve of SORG-MLA is above the other two competing models with threshold probabilities from 0.1 to 0.8. <h3>Conclusion</h3> Laboratory data are of prognostic significance in survival prediction after RT for SM. Machine learning-based model SORG-MLA outperforms statistical regression-based model METSSS and NESMS in predicting 90-day and 1-year survival after RT for patients with SM.