# Identification of Glioma Phenotypic Subtypes From Multimodal MRI Data Using Hierarchical Multi‐Kernel Learning

## TL;DR

This study uses MRI data and machine learning to identify glioma subtypes and their associated pathways, helping guide treatment decisions.

## Contribution

A hierarchical multi-kernel learning approach was used to identify glioma subtypes and predict IDH genotype from MRI data.

## Key findings

- Two glioma phenotypic subtypes were identified with distinct survival outcomes and pathway activities.
- The GA-KPLS model achieved an AUC of 0.819 for predicting IDH genotype from MRI data.
- High-risk gliomas showed activation of JAK–STAT and TGF-β pathways, while low-risk gliomas showed Hypoxia and p53 pathway activation.

## Abstract

Gliomas are the most common primary brain tumors, exhibiting significant phenotypic variability even within the same grade. Identifying glioma subtypes through non‐invasive methods could improve patient management.

In this study, we applied hierarchical multi‐kernel learning to identify glioma phenotypic subtypes using MRI data (T1CE and T2FLAIR) from the First Hospital of Shanxi Medical University (FHSXMU) and Shanxi Provincial People's Hospital (SPPH) (n = 246). We further validated our findings using an independent dataset of similar tumor characteristics from The Cancer Genome Atlas/The Cancer Imaging Archive (TCGA/TCIA) (n = 144). Additionally, we analyzed pathway activity across glioma subtypes from the TCGA/TCIA dataset and employed five machine learning models, namely kernel partial least squares with the genetic algorithm (GA‐KPLS), random forest, the least absolute shrinkage and selection operator, K‐Nearest Neighbor, and Naïve Bayes, to predict isocitrate dehydrogenase (IDH) genotype from the FHSXMU/SPPH dataset.

We identified 2 glioma phenotypic subtypes, high‐risk and low‐risk groups. These groups showed significant differences in overall survival (p < 0.05) and were associated with distinct signaling pathways. The JAK–STAT and TGF‐β pathways were activated in the high‐risk group, while the Hypoxia and p53 pathways were activated in the low‐risk group. Among the machine learning models, the GA‐KPLS model demonstrated the highest predictive performance for the IDH genotype, achieving an area under the curve of 0.819.

Our study provides a non‐invasive method to identify glioma phenotypic subtypes, reveal distinct signaling pathways, and define therapeutically homogeneous patient subgroups that could guide targeted therapy.

Our study identified glioma phenotypic subtypes based on T1CE and T2FLAIR radiomic features, along with associated pathway activities, providing real‐time insights into tumor dynamics to guide preoperative clinical decision‐making and improve patient management. Additionally, the GA‐KPLS‐based predictive model effectively assessed IDH genotype, enabling preoperative evaluation to inform clinical decisions.