Integrating CT radiomics, transcriptomic module features, and clinical variables using a random forest model for gastric cancer survival-status classification

Gastric cancer remains a common gastrointestinal malignancy and a major contributor to cancer-related mortality worldwide. Prognostic prediction still relies heavily on clinicopathological factors. Multimodal machine learning may integrate quantitative imaging, transcriptomic, and clinical information for survival-status classification. CT imaging data, transcriptomic profiles, and clinical records from 46 matched gastric cancer patients were obtained from TCGA-STAD/TCIA resources. A total of 116 CT radiomic features were extracted using PyRadiomics, and 11 features were retained by least absolute shrinkage and selection operator-based L1-regularized logistic regression. Transcriptomic data were normalized by Z-score transformation, and Weighted Gene Co-expression Network Analysis was used to derive module eigengenes. The cohort was organized into a training/internal-validation set ( n = 37) and a held-out evaluation set ( n = 9). Stratified five-fold cross-validation was performed within the training/internal-validation set, and the held-out evaluation set was used for the displayed confusion matrix and final exploratory performance reporting. Eleven CT radiomic features were retained for multimodal modeling. Weighted Gene Co-expression Network Analysis identified seven transcriptomic modules, and their module eigengenes were used as low-dimensional transcriptomic predictors. The random forest model achieved an area under the receiver operating characteristic curve of 0.85 (95% confidence interval, 0.68–0.97) in stratified five-fold cross-validation. In the held-out evaluation set, the model achieved an accuracy of 0.78, with precision and recall for the Survival class of 0.80 and 0.80, respectively, and a weighted F1 score of 0.78. This study provides an exploratory random forest-based multimodal model for gastric cancer survival-status classification using CT radiomics, transcriptomic module features, and clinical variables. Larger independent cohorts are required for external validation and assessment of model generalizability. Not applicable (this study involved retrospective analysis of publicly available data and did not report results of a health care intervention on human participants).