Click here to

Session: Therapy: Outcome Modeling and Assessment II [Return to Session]

Prediction of Glioblastoma Patient’s Survival After Radiation Therapy with Random Survival Forest Model

Y Kim1*, KW Kim2, H Yoon2, W Sung1, (1) The Catholic University of Korea, Seoul, ,KR (2) Yonsei University College of Medicine, Seoul, ,KR


MO-IePD-TRACK 5-2 (Monday, 7/26/2021) 3:00 PM - 3:30 PM [Eastern Time (GMT-4)]

Purpose: To predict the outcome of radiation therapy (RT) in patients with glioblastoma from parameters before and after RT and to identify predictive variables using a machine learning.

Methods: We assessed 336 glioblastoma patients treated between 2006 and 2017. A random survival forest was trained to predict survival from baseline values of total 16 parameters including tumor, demographic, immunologic, and treatment factors. The feature importance was determined using the amounts of increased predictive power by adding each parameter.

Results: The trained model achieved concordance index (c-index) of 0.628 in cross-validation, comparable to the one (0.629) in the similar published study with cox proportional hazard regression. The importance analysis identified methylation of the MGMT gene promoter, the extent of resection, and sub-ventricular zone involvement as the most important variables (permutation importance = 0.082, 0.034, and 0.021 respectively), followed by gross target volume (0.018), age (0.008), and acute severe lymphopenia (0.004, lymphocyte counts of < 500 μL). More number of variables did not necessarily guarantee the predictive power of the trained model. The model achieved the maximal concordance index (0.658) by including the most important 3 or 4 parameters.

Conclusion: Random survival forests were an explainable machine-learning approach for prediction of an entire patient’s survival curve. Only three or four baseline parameters were necessary to achieve comparable predictive power of the trained model. Those parameters were the methylation of MGMT gene promoter, extent of resection, sub-ventricular zone involvement, or gross target volume, for the glioblastoma patients.

Funding Support, Disclosures, and Conflict of Interest: The authors wish to acknowledge the financial support of the Catholic Medical Center Research Foundation made in the program year of 2020.



    X Rays, Tumor Control, Modeling


    TH- Response Assessment: Modeling: Machine Learning

    Contact Email