Training and validation cohort analysis for predicting radiation therapy response in human glioblastoma.

Abstract

e13018 Background: Glioblastomas comprise 50% of primary brain tumors and are the most fatal, with a median survival time of 14 months despite aggressive treatment. Due to disease heterogeneity and the presence of subclinical disease, there is a large variability in treatment response among patients, making it difficult to assess treatment efficacy and confounding clinical decision making. Methods: A patient-specific, mathematical model of glioblastoma has been shown to predict untreated growth as well as the effect of radiation therapy. We expand on the technique presented in Rockne 2010 to determine patient-specific radiosensitivity and post therapy radial growth using only pretreatment data. From a potential cohort of 44 patients we randomly chose a training set of 30 to compute the relationship for the radiosensitivity parameter (alpha) versus the net proliferation rate. We then used that relationship to compute an individualized alpha for the remaining 14 test patients based only on pretreatment information. For each of the test patients we compared the observed T1Gd visible radius at the first post radiotherapy timepoint with the simulated prediction. Results: Half of the patients demonstrated average absolute differences between the observed and simulated post radiotherapy T1Gd radii of less than 5 mm, within the measurement error associated with contouring the tumor on MRI, with a maximum difference of 2.5 cm. The largest errors were observed in patients with significant resection. Conclusions: This patient-specific, mathematical model of glioma growth and response to radiotherapy can potentially predict radioresponse in vivo, prior to the commencement of therapy. As a clinical tool, this can have prognostic applications and separate responders from non-responders before deciding to treat with radiotherapy. The larger error for patients with more extensive resections is not unexpected given the difficulties in contouring and assessing tumor burden post surgery. This can be improved by including other imaging modalities such as pre-gadolinium T1 and improving the quantification of extent of resection.