Multi-path convolutional neural network for glioblastoma survival group prediction with point mutations and demographic features

Abstract

Glioblastoma multiforme (GBM) is the most common and aggressive brain cancer with a median survival rate of 15 months. It is well-established that age is a strong independent predictor of GBM survival outcome. There is accumulating evidence that single nucleotide polymorphisms (SNPs) in the IDH1 gene influences GBM survival time. We propose a new multi-path convolutional neural network that combines SNPs, age, age groups, and gender to predict survival groups with a one-year threshold. We obtained GBM SNP and demographic data from The Cancer Genome Atlas. We compare our multipath CNN with a support vector machine (SVM) and random forest. We randomly held out 10% of the samples as a test set, and employed 10-fold cross-validation for hyperparameter tuning in the remaining 90%. We then fit a model with optimal hyperparameters and predict the test set. In the combined SNP and demographic features, our proposed multi-path model achieved 67% accuracy in the test set compared to SVM accuracy of 60% and random forest accuracy of 47%. In the 10-fold cross-validation, our model predicted the two survival groups with 63% mean balanced accuracy while SVM and random forest attained 56% and 49% mean balanced accuracy. We evaluated the predictive ability in combined SNP and demographic data versus each data source alone for our proposed CNN, SVM, and random forest. The highest achieved accuracy for SNP data only in the test data set is 60% with our single-path CNN. The top accuracy in the test data set for demographic features alone attained is 60% by SVM and our single-path neural network.