The class label of some multi-class machine-learning problems may be hierarchical. A well-known example is lower-grade glioma, which can be clinically classified first based on IDH mutation and then according to chromosome 1p/19q codeletion. Hi-Class is a newer supervised framework that leverages the hierarchical information within class labels for more accurate and efficient prediction. This study demonstrates the application of Hi-Class to predict the established glioma subtypes using cytosine-phosphate-guanine island methylation in two independent, real-world datasets (n=507 and n=122). As an exemplar of reproducibility, hold-out and cross-validation experiments, with or without Hi-Class, were performed in both cohorts. To exemplify generalizability, both Hi-Class and flat model versions were trained on one dataset and applied to another without between-cohort data processing. The strong prior knowledge and the observed classification performance make this task suitable for demonstrating and teaching specialized use cases of multi-class predictive analytics.

Author: Youdinghuan Chen

Published in: Canada International Conference on Education, 2024

• Date of Conference: 23-25 July, 2024
• DOI: 10.20533/CICE.2024.0082
• Electronic ISBN: 978-1-913572-65-5
• Conference Location: Toronto Metropolitan University, Toronto, Canada