Document Type


Publication Date



Introduction: The role of surgical resection for the treatment of glioblastoma multiforme is well established. Survival analysis after resective surgery in the literature comprises mostly of traditional statistical models. Machine learning models offer powerful predictive and analytical capability for varied datasets and offer improved generalizability and scalability. We analyzed survival data of patients with glioblastoma with various machine learning algorithms and compared it to binary logistic regression.

Methods: We retrospectively identified cases of glioblastoma treated with surgical resection at our institution from 2012-2018. Feature scaling and one-hot encoding was used to better fit the models to the data and used the formula X’ = (X – Xmin)/(Xmax – Xmin). Feature selection was performed using chi-squared analysis (features with p

Results: 582 patients fit the inclusion criteria and were used to build these models. 6-month mortality was 43.13%. Accuracy scores (AUC) for models used were 0.670 (logistic regression), 0.704 (Random Forest), 0.585 (Support Vector Machine), 0.560 (Naïve Bayes), 0.650 (XG Boost), 0.585 (Stochastic Gradient Descent Classifier), and 0.740 (Neural Network). 5-fold cross validation was used to ensure generalizability to an independent dataset.

Conclusion: Machine learning methods for prediction of six-month survival for glioblastoma are promising analytical tools that we show can approach or exceed the accuracy of traditional logistic regression, particularly neural networks and the random forest algorithm. Improved prediction of 6-month survival using machine learning offers increased capabilities for patient education, adjuvant chemotherapy or radiation planning, and post-operative counseling, while maintaining increased adaptability and generalizability compared to regression models.