PATH-22. COMPARISON OF SUPERVISED CLASSIFICATION METHODS FOR CENTRAL NERVOUS SYSTEM TUMORS BASED ON DNA-METHYLATION

Abstract

Classification of brain tumors using methylation profiling is an important diagnostic advance, reducing subjectivity and improving interpretability of clinical outcome data. Despite the recognized value of methylation profiling in the clinical laboratory, the performance characteristics of different supervised classification models has not been directly compared. We developed 3 methods using methylation profiles to classify CNS tumors: an exact bootstrap k-nearest neighbor (kNN), a multi-layer perceptron neural net (NN), and a random forest classifier (RF). We trained these methods on the publicly available CNS tumor reference cohort (GSE90496) with 2,801 profiles and 91 classes. We evaluated the performance of these methods by leave-out-25% cross-validation. The relative performance of these methods were evaluated in terms of accuracy, precision, and recall for class or class family. The kNN, RF, and NN classifier had an estimate error rate of 10.74%, 4.01%, and 1.89%, respectively for class prediction and an error rate for family prediction of 5.97%, 0.90%, and 0.6%, respectively. At perfect recall for class assignment, the RF and kNN had a precision of 0.96 and 0.89 while the NN reached 0.98. For family assignment, the precision for the three classifiers was almost 1.0 with recall of nearly 0.8. At the recall rate of 1.0, the precision dropped to 0.94, 0.991 and 0.994 for kNN, RF, and NN, respectively. Overall, the NN showed improved performance metrics compared to the kNN and RF in CNS tumor classification for both class and class family assignment.