Machine-Learning to Predict Survival in Glioblastoma based on Growth Kinetics
Corbin Rayfield, Mayo Clinic, Scottsdale, AZ; Scott Whitemire, Mayo Clinic, Scottsdale, AZ; Andrea Hawkins-Daarud, Mayo Clinic, Scottsdale, AZ; Kristin Swanson, Mayo Clinic, Scottsdale, AZ
High-grade gliomas, commonly known as glioblastoma multiforme (GBM), exhibit linear radial expansion. In untreated patients, these growth kinetics are believed to be constant. Not much is known about the changes in growth kinetics that different treatment regimes induce. In this study we sought to compare pre-treatment growth kinetics to post-treatment growth kinetics; investigate correlations between pre-, during, and post-treatment growth kinetics survival, and elucidate measures of response to treatment that define post-treatment growth characteristics. Additionally, we present a novel machine-learning algorithm to predict which survival cohort a patient will fall into utilizing growth velocities.
Patient information and MRIs were obtained from 101 newly diagnosed patients with glioblastoma from Northwestern University, University of Washington, Columbia University, and University of California, Los Angeles who had at least 2 T1Gd images in the pre-treatment or radiation period and analyzed retrospectively. The tumor growth kinetics were calculated by selecting 2 imaging events, separated by at least 4 days, and calculating the rate of growth according to:
We defined a pre-treatment velocity, radiation velocity, post-treatment velocity, and time to nadir as kinetics to investigate. The time to nadir represented the amount of time between the end of radiation and the smallest tumor radius.
The pre-treatment and radiation velocity did not have a significant correlation with overall survival. Additionally, the pre-treatment velocity did not correlate with the post-treatment velocity. The post-treatment velocity and the time to nadir had a significant correlation with overall survival (likelihood ratio test= 14.9, p-value= 0.001; likelihood ratio test= 36.5, p=1.5e-09, respectively). These kinetics remained significant on multivariate analysis (p-value=3.73e-05; p-value=3.55e-07)
For survival predictions, patients were grouped according to quartile of survival. Group 1 contained patients with survival less than the 33rd percentile (13.63 months), Group 2 contained patients with survival between the 33rd and 66th percentile (23.53 months), and Group 3 contained patients with survival greater than the 66th percentile. A model based on recursive partitioning analysis was generated on a training set of 85 patients and tested on the remaining 16 patients. The test cohort was then iterated such that 10 new test cohorts were generated. The accuracy of prediction was the median accuracy of all 10 tests.
We found that utilizing time to nadir, post-treatment velocity, age, and radiation velocity, the survival cohort could be accurately predicted 70% of the time. Additionally, this prediction had an area under the curve of the receiver-operating characteristic of 80%. Finally, we found that we could predict extreme survival (patients with survival > 3 years) 88% of the time.Format: Poster