Multimodal data integration using machine learning improves prognosis risk stratification of glioblastoma patients

The prognosis for glioblastoma(GBM) remains dismal. An accurate assessment of prognostic stratification is crucial for guiding clinical decisions. We aimed to establish and validate a multimodal model for individualized and accurate survival risk stratification in patients with GBM. We collected a multimodal dataset comprising 194 consecutively enrolled IDH-wildtype GBM patients. First, we performed analysis of clinical features using Cox regression, resulting in a clinical signature model. Second, we use ResNet101, DenseNet121, and Inception_v3 deep learning neural networks, extraction radiomics features, formed deep learning radiomics signature (DL-Rad_Sig), we also extracted pathomics features based on H&E-stained whole slide images (WSIs), resulting in the pathomics signature (Path_Sig). The full cohort was split into training (136 patients) and test (58 patients) sets via stratified random sampling based on 1-year overall survival status. LASSO-Cox regression with 10-fold cross-validation within training set was adopted for feature screening. Finally, a multimodal model was constructed to predict overall survival (OS) of GBM patients. The clinical features performed suboptimal in prognosis analysis. DL-Rad_Sig model, the c-index values of 0.852 and 0.721 for the train and test cohort. Six pathomics features were used develop Path_Sig modeling, resulting in c-index values of 0.821 and 0.762 in the train and test cohort. And we compared the fusion of the three signatures with each individual signature. We found that the combined model exhibited a better C-index, the C-index of 0.887 and 0.791 for OS in train and test cohort. Kaplan-Meier survival curves of the combined model nomogram show the OS rate difference was statistically significant between the high- and low-risk patients. The ROC curve of the combined model nomogram at predicts 1-, 2-, and 3-year OS produced AUC of 0.871, 0.725, and 0.833, respectively. Our internally validated multimodal nomogram integrating clinical, deep radiomics and weakly-supervised pathomics signatures achieves favorable OS risk stratification for single-center IDH-wildtype glioblastoma patients, though external multi-center validation is still required to verify generalizability. Not applicable. Both deep learning radiomics and pathomics models achieve effective prognostic stratification for GBM patients. Pathomics features extracted from H&E-stained WSIs are significantly associated with clinical outcomes in GBM. The multimodal fusion nomogram serves as a reliable and non-invasive tool for prognostic evaluation of GBM patients. This study only performed single-center internal validation without independent external cohorts, and patch-level pathomics models show obvious training-set overfitting which needs cautious interpretation.