Construction and validation of a prognostic nomogram model integrating machine learning-pathomics and clinical features in IDH-wildtype glioblastoma.

BACKGROUND: Novel diagnostic criteria for glioblastoma (GBM) in the 2021 WHO classification emphasize the importance of integrating pathological and molecular features. Pathomics, which involves the extraction of digital pathology data, is gaining significant interest in the field of tumor research. This study aimed to construct and validate a nomogram based on machine-learning pathomics for patients with GBM. METHODS: We extracted pathomic features from hematoxylin and eosin (H&E)-stained images of GBM from the Department of Neurosurgery of Nanfang Hospital (n = 125), Department of Neurosurgery of Zhangzhou Affiliated Hospital of Fujian Medical University (n = 96), and The Cancer Genome Atlas (n = 104) using CellProfiler. We then constructed a machine learning-pathomics risk score (PRS) model using the LASSO (least absolute shrinkage and selection operator)-Cox regression method. Clinical data including sex, age, preoperative Karnofsky performance status (KPS), extent of resection, subventricular zone (SVZ) association, and O6-methylguanine-DNA methyltransferase (MGMT) promotor methylation status, were also obtained. Differentially expressed gene analysis, gene ontology analysis, and immunohistochemical staining were utilized to establish a link between PRS and GBM molecules. We subsequently constructed a nomogram model integrating PRS with other independent clinical risk factors and was then validated externally. RESULTS: Ten pathomics features were identified using the PRS model. An association between the PRS, tumor location, and molecular characteristics was observed. Notably, the PRS is related to the extracellular matrix, including type 1 and type 6 collagen. Patients with a low PRS, but not those with a high PRS, significantly benefited from supramaximal resection. Moreover, the combination of the PRS, KPS, extent of resection collectively formed a novel prognostic nomogram model with high accuracy. CONCLUSIONS: This novel prognostic nomogram model integrating machine learning pathomics and clinical features for GBM patients, is available as free online software at https://yaomin.shinyapps.io/GBM_Pathomics_Nomogram_NFH/ .