Predictive Modeling of Brain Metastasis Velocity in a Validated Multi-Institution Cohort

Abstract

Brain metastasis velocity (BMV), a recently described prognostic metric for patients receiving stereotactic radiosurgery (SRS), describes the number of new brain metastases per year, and stratified patients based on low (<4), intermediate (4-13), and high (> 13) BMV. This grouping has also been correlated with overall survival, neurologic death, and need for salvage WBRT. This finding after both initial and repeat SRS has been validated independently by multiple groups internationally. At this time, there are no models that are able to predict patient BMV at the time of upfront SRS. This would allow for more appropriate patient risk stratification and clinical management. A total of 2546 patients treated with upfront stereotactic radiosurgery (SRS) at nine institutions were included for this potential analysis, and 1647 had appropriate follow up to calculate BMV. A representative validation set of 163 patients were excluded from the total data set to generate predictive models. Patient characteristics that were investigated for potential inclusion in this model were gender, age, primary histology, RPA classification, number of metastases at initial SRS, systemic disease status, disease burden, and margin dose, which is a surrogate for metastasis size. In the high BMV model the number of metastases at first SRS, primary histology, and age were significant predictors. In the low BMV model metastases number, primary, disease burden, gender, and margin dose were significant. Using a data and decision management software, a logistic regression modeling technique with stepwise selection with an entry level and stay level of 0.15 was used to independently generate models for both low BMV and high BMV outcomes. The cut point to determine the binary outcome of each individual model was based on the midpoint between the different predicted probability means. The mean of the two groups’ predicted probabilities was statistically significantly difference in a t-test with a p value of <0.0001 in both high and low BMV models. Patients included in this analysis had a median age of 60, median number of brain metastases at initial SRS of 2, median follow up of 12.69 months, and median overall survival of 14.17 months. In the model dataset the overall accuracy of the model was 74% in the high BMV and 67% in the low BMV model. The high BMV model had a sensitivity of 0.4960 and a specificity of 0.7880. The low BMV model had a sensitivity of 0.7268 and a specificity of 0.5598. The predicted probabilities from the validation set were able to accurately group 70 (70%) patients with low BMV and 14 (52%) patients with high BMV. We have generated a rigorous statistical model that is able to reliably predict the potential that a patient will fall within the high or low BMV risk group.