GlioPredictor: A Deep Learning Model for Identification of High-Risk Low-Grade Glioma toward Adjuvant Treatment Planning.

PURPOSE/OBJECTIVE(S): High-risk low-grade glioma (LGG) patients are recommended to undergo adjuvant radiotherapy whereas watchful waiting is recommended for low-risk LGG patients per the latest NCCN guidelines. Based on Radiation Therapy Oncology Group (RTOG) 9802, high-risk features include age >40 or subtotal resection (STR). However, in the era of molecular-based classification for tumors of central neural system, current risk classification criteria based on gross disease and patient demographics may be outdated. Here, we aim to develop a molecular-based glioma risk classification system (GlioPredictor) that could potentially facilitate identification of high-risk LGG patients.

MATERIALS/METHODS: A total of 507 LGG cases from The Cancer Genome Atlas-low grade glioma (TCGA-LGG), and 1,309 cases from AACR GENIE v13.0 datasets were studied for genetic disparities between IDH1-wildtype and mutated cohorts, and varying age groups. Through a feature selection technique using genomic profiling and correlation analyses, features such as mutation status, copy number variations (CNVs), among other clinicopathologic features prognostic of IDH1 mutation status were selected as potential inputs to train an artificial neural networks (ANNs) that could predict IDH1 mutation status. Model performance was assessed using the area under the receiver operating characteristic curve (AUC). Memorial Sloan Kettering (MSK) dataset (n = 404) for LGG was used to cross-validate the trained ANN. The optimized ANN model has 6 layers with 6 input nodes, 20 hidden nodes, and a binary output layer. The weights and biases of the hidden layers of the best-performing model were retrieved and reconstructed to yield the GlioPredictor score-the predicted risk of progression for IDH1-wildtype LGG.

RESULTS: Over 81% of glioma patients age less than 40 have IDH1 mutation, as compared with 31% in those age above 60. Using age > 40 as a cutoff failed to identify high-risk IDH1-mutant LGG with early progression. IDH1 mutation is associated with decreased CNVs of EGFR (21 % vs. 3%), CDKN2A (20% vs. 6%) and PTEN (14% vs. 1.7%), and increased percentage of mutations for TP53 (15% vs. 63%), and ATRX (10% vs. 54%) (p<0.001). Using these molecular features, along with the patient's age, an ANN model with 6 layers and 20 hidden nodes can predict IDH1 mutation status with over 90% accuracy and AUC score over 0.91.

CONCLUSION: We have developed an ANN model that is capable of learning the prognostic features of LGG associated with an IDH1-mutated LGG cohort and using the features to predict high-risk patients from the IDH1-wildtype cohort. This ANN model facilitates the selection of LGG patients who could benefit from immediate adjuvant radiotherapy. Future work includes the integration of image features to improve the prediction performance of the GlioPredictor system.