Clinical Features at Presentation for Glioblastoma Patients Impact Survival Predictions in a Machine Learning Model

Abstract

Abstract Aims To generate an accurate prediction model for greater than median survival using Random Forest machine learning analysis and to compare the model to a traditional logistic regression analysis model on the same Glioblastoma Dataset. Method In this single centre retrospective cohort study, all patients with histologically diagnosed primary GB from October 2014 to April 2019 were included (n=466). Machine learning algorithms encompassing multiple logistic regression and a Random Forest, Gini index-based decision tree model with 100,000 trees were used. 17 clinical, molecular and treatment specific binarily categorised variables were used. The dataset was split 70:30 into training and validating sets. Results The dataset contained 466 patients. 326 patients made up the training set and 140 the validation set. The Random Forest model’s accuracy for predicting 18-month survival was 86.4% compared to the Logistic Regression model’s accuracy of 85.7%. The top 5 factors that the Random Forest model used to predict survival over 18 months were; mean MGMT status >10%, if the patient underwent gross total resection, whether the patient had adjuvant temozolomide, whether the patient had a neurological deficit on presentation, and the sex of the patient. Conclusion Machine learning can be applied in the context of GB prognostic modelling. The models show that as well as the known factors that affect GB survival, the presenting symptom may also have an impact on prognostication.