Development of a web-based tool for estimating individualized survival curves in glioblastoma using clinical, mRNA, and tumor microenvironment features with fusion techniques.

Abstract

Objective: Glioblastoma (GBM), one of the most common brain tumors, is known for its low survival rates and poor treatment responses. This study aims to create a robust predictive model that integrates multiple feature types, including clinical data, RNA expression, and tumor microenvironment data, using fusion techniques to enhance model performance.

Methods: We obtained data from the SEER database to assess the impact of nine demographic and clinical features on the survival of 58,495 GBM patients and built predictive machine learning models. Additionally, mRNA expression data from 600 GBM patients from TCGA, CGGA, and GEO were analyzed. We used Cox regression and LASSO to create a gene signature, which was compared against 13 published signatures for accuracy. Twenty-one machine learning models were applied to predict survival at multiple time points. Finally, we integrated multiple feature types using fusion techniques and developed a Shiny app to provide survival predictions for GBM patients.

Results: Using the SEER database, we constructed machine learning models based on nine clinical variables: age, gender, marital status, race, tumor site, laterality, surgery, chemotherapy, and radiation therapy. The best-performing model achieved AUC values of 0.775, 0.728, 0.692, and 0.683 for predicting survival at 6, 12, 18, and 24 months in the testing cohort. In the merged TCGA, CGGA, and GEO cohorts, we identified 11 genes to develop predictive models. These 11 genes outperformed 13 other published gene signatures in predicting the prognosis of GBM. When incorporating mRNA features, tumor microenvironment features, and clinical variables into the fusion models, the AUC values for predicting survival at 6, 12, 18, and 24 months were 0.641, 0.624, 0.655, and 0.637, respectively. A user-friendly tool for predicting the survival curve of individual GBM patients is available at https://zzubioinfo.shinyapps.io/mlGBM/ .

Conclusions: Our study provides a web-based tool that includes two modules: one for predicting survival curves using only clinical variables, and another that integrates multiple feature types for more comprehensive predictions.