Multiclass machine learning models for molecular subtype identification of pediatric low-grade glioma using bi-institutional MRIs for precision medicine.
Authors
Affiliations (18)
Affiliations (18)
- Department of Diagnostic & Interventional Imaging, The Hospital for Sick Children (SickKids), Toronto, ON, Canada.
- Neurosciences & Mental Health Research Program, SickKids Research Institute, Toronto, ON, Canada.
- Institute of Medical Science, University of Toronto, Toronto, ON, Canada.
- Vector Institute, Toronto, ON, Canada.
- Department of Medical Imaging, University of Toronto, Toronto, ON, Canada.
- Department of Diagnostic and Interventional Neuroradiology, University Hospital Augsburg, Augsburg, Germany.
- Division of Neurosurgery, The Hospital for Sick Children (SickKids), Toronto, ON, Canada.
- Department of Radiology, Phoenix Children's Hospital, Phoenix, AZ, USA.
- Department of Neurosurgery, Stanford School of Medicine, Stanford, CA, USA.
- Department of Paediatric Laboratory Medicine, Division of Pathology, The Hospital for Sick Children, University of Toronto, Toronto, ON, Canada.
- Department of Neurooncology, The Hospital for Sick Children, Toronto, ON, Canada.
- Department of Diagnostic & Interventional Imaging, The Hospital for Sick Children (SickKids), Toronto, ON, Canada. [email protected].
- Neurosciences & Mental Health Research Program, SickKids Research Institute, Toronto, ON, Canada. [email protected].
- Institute of Medical Science, University of Toronto, Toronto, ON, Canada. [email protected].
- Vector Institute, Toronto, ON, Canada. [email protected].
- Department of Medical Imaging, University of Toronto, Toronto, ON, Canada. [email protected].
- Department of Computer Science, University of Toronto, Toronto, ON, Canada. [email protected].
- Department of Mechanical and Industrial Engineering, University of Toronto, Toronto, ON, Canada. [email protected].
Abstract
Pediatric Low-Grade Glioma (pLGG) is the most common pediatric brain tumor, and radiomics-based machine learning (ML) models have shown promise in identifying BRAF fusion and BRAF p.V600E mutation. This bicentric retrospective study included 495 children diagnosed between 1999 and 2023. The local hospital dataset comprised Magnetic Resonance Imaging (MRI) scans of patients with BRAF fusion (n = 190), BRAF p.V600E mutation (n = 95), FGFR1 (n = 25), and other molecular subtypes (n = 144), while an external dataset included BRAF fusion (n = 32) and BRAF p.V600E mutation (n = 9) cases. Radiomics features were extracted from Fluid-Attenuated Inversion Recovery images, and Random Forest classifiers were trained using Monte Carlo data splits and leave-one-out validation. The best-performing model achieved an average one-vs-the-rest area under receiver operating characteristic curve of 0.819 (95% confidence interval [0.791, 0.848]). This study highlights the potential of radiomics-based ML models for molecular subtype differentiation in pLGG, with per-patient predictions enabling outlier identification and subgroup performance evaluation.