Construction of a prognostic model for glioblastoma through the integration of conventional pathology and clinical features using deep learning methods
YANG Haohui
LIAN Jiabian
XU Tao
LUAN Zhonghua
AO Yongfang
ZHANG Yuhao
HAN Yudong
ZHU Jiabao
Abstract：Objective To construct and validate a robust prognostic model based solely on the integration of conventional hematoxylin-eosin-stained whole slide imaging(WSI)and clinical features,with the goal of enhancing its clinical applicability.Methods A total of 1,136 WSIs and clinical data from 300 patients in The Cancer Genome Atlas Glioblastoma Multiforme(TCGA-GBM)dataset were retrospectively analyzed.The data were randomly divided into a training set(210 patients)and a test set(90 patients)at a 7:3 ratio.Using a weakly-supervised deep learning framework,prognostic-related features were extracted from over 3.4 million image patches based on the median survival time(15 months).The patch likelihood histogram(PLH)was systematically compared and selected as the patient-level feature aggregation strategy.After integrating clinical variables,13 core features were selected through LASSO regression.Eight machine learning survival models,including random survival forest(RSF)and the LASSO-Cox model,were used for ensemble modeling and evaluation.Results The RSF model demonstrated the best performance in the test set,with time-dependent area under the curve(AUC)values of 0.829,0.887,0.968,0.911,0.841,and 0.848 at 6,12,15,18,24,and 36 months,respectively,showing stable discriminatory performance from the early to medium-long term,with a concordance index(C-index)consistently above 0.782.The model demonstrated good calibration performance,and decision curve analysis revealed significant clinical net benefits across a wide range of threshold probabilities.The survival difference between the high-risk and low-risk groups,divided based on the model's risk score,was highly significant hazard ratio(HR)=5.184,95%confidence interval(CI):4.003-6.712,P<0.0001).Visual analysis indicated that the morphological features identified by the model were qualitatively consistent with typical regions of poor prognosis,such as necrosis and microvascular proliferation.Conclusions The"pathomics-clinical feature"prognostic model for glioblastoma(GBM)patients demonstrates excellent predictive accuracy,robustness,and clinical practicality.It is particularly adept at identifying early high-risk patients,providing an artificial intelligence tool based on conventional data with the potential for visual interpretation and easy dissemination for individualized postoperative prognostic assessment in GBM patients.
Keywords：gliomaglioblastomaprognostic predictionpathomicsclinical featuresdeep learning
Publication Date：2026-02-20
Online Publishing Date：2026-03-18(First online date of this platform, not the publication date of the document)
Pages：9( 73-81 )