87 Views

A machine learning model of radiomic, molecular, and clinical determinants of aphasia in diffuse glioma

Friday, May 3, 2024

Presenting Author(s)

Abraham Dada, BA

Medical Student
Department of Neurological Surgery, University of California, San Francisco

Co-Author(s)

Alexander A. Aabedi, MD

Resident Physician
Department of Neurological Surgery, University of California, San Francisco
San Francisco, California, United States
EA

EunSeon Ahn, PhD

Graduate Student
University of Michigan, United States
DB

David Brang, PhD

Assistant Professor
University of Michigan, United States
AC

Abrar Choudhury, MD, PhD

Resident Physician
Department of Neurological Surgery, University of California, San Francisco, San Francisco, CA, USA, United States
Shawn Hervey-Jumper, MD

Associate Professor
Department of Neurological Surgery, University of California, San Francisco
San Francisco, California, United States
JK

Jasleen Kaur, BS

Clinical Researcher
Department of Neurological Surgery, University of California, San Francisco, United States
SK

Saritha Krishna, PhD

Staff Scientist
Department of Neurological Surgery, University of California, San Francisco, United States
Thomas A. Wozny, MD

Resident Physician
Department of Neurological Surgery, University of California, San Francisco
San Francisco, California, United States

Introduction: While gliomas have been likened to ablative lesions, our recent work suggests that glioma-infiltrated cortex still engage in task-specific cognitive processing but with reduced information encoding. Tumor volume and local mass effects have been linked to cognitive deficits, but the role of imaging features and tumor-specific mutations in mediating these effects remains unknown.

Methods: Eighty-eight UCSF patients with newly diagnosed WHO 2-4 diffuse glioma, Next Gen Sequencing (NGS), and complete Quick Aphasia Battery (QAB) language assessments were included. Radiographic features were extracted from MRI sequences. Support-vector regression voxel-lesion symptom mapping (SVR-LSM) revealed an anatomic region of interest (ROI) associated with lower QAB scores. A random forest model identified features (NGS, radiographic, and clinical) most predictive of aphasia. Top genes driving the model were investigated through single cell RNA sequencing.

Results: SVR-LSM analysis revealed full lesions within a cluster of 61,745 voxels in the left superior temporal gyrus were predictive of lower QAB scores. The top three features driving the model were: number of T1 Post-Gad voxels within the ROI, number of full lesion voxels within the ROI, and T2-FLAIR voxels within the ROI. Genetic mutations driving the model included ARID2, polysomy 7 monosomy 10, PLCB4, LZTR1, MDM4, NF1 and CDKN2A/B co-deletion. Variations in ARID2, known for encoding chromatin remodeling proteins, were associated with worse QAB scores (6.56 vs. 9.14 p=0.0055). Multivariate analysis confirmed that ARID2 variant status (p=0.02) and T1-Post Gad voxels overlapping with the ROI (p < 0.001) significantly predict lower QAB scores, controlling for confounding clinical variables. In single-cell sequencing analysis, ARID2 was predominately expressed in diving tumor cells.

Conclusion : Glioma patients with ARID2 variants exhibit impaired language function, with a linear relationship between T1 Post-Gad overlap with neuroantomic language centers and impairment severity. This translational study implicates a novel target, ARID2, in altering language function in diffuse gliomas.
.