A radiomics-driven machine learning model for predicting bladder cancer prognosis identifies genes associated with radiomic features.

background and purposeBladder cancer (BC) exhibits pronounced tumor heterogeneity, which poses a persistent clinical challenge for accurate prognostic prediction and monitoring. This study aims to develop and validate an interpretable machine learning model that leverages radiomics features to predict BC outcomes, while also investigating associated genetic signatures and their relationship with tumor immunity.

methodsRNA-seq data were acquired from The Cancer Genome Atlas (TCGA) and the Gene Expression Omnibus (GEO). Clinicopathological parameters and radiomics features for bladder cancer (BC) cases were obtained from TCGA and The Cancer Imaging Archive (TCIA). Differentially expressed genes (DEGs) were identified from both TCGA and GEO datasets. Key radiomics features were selected through univariate Cox regression followed by LASSO-Cox regression, while independent clinicopathological prognostic factors were determined using univariate and multivariate Cox regression analyses. A support vector machine (SVM)-based prognostic model was developed by integrating clinicopathological and radiomics features. The performance of this model was evaluated against three other machine learning models using ROC curves, accuracy, precision, and F1-score. Feature importance scores were computed, and model interpretability was enhanced using SHapley Additive exPlanations (SHAP). Based on the feature importance coefficients derived from the SVM model, BC patients were classified into distinct risk groups. Subsequently, radiomics-associated DEGs were analyzed to pinpoint key genes linked to imaging phenotypes. These genes were subjected to functional annotation via Gene Ontology (GO), pathway enrichment analysis using KEGG, gene set enrichment analysis (GSEA), and immune infiltration analysis to investigate potential patterns of immune cell enrichment.

resultsAfter rigorous screening, 91 bladder cancer patients with high-quality enhanced CT images, complete clinicopathological data, and RNA sequencing profiles were included in the study. These patients were randomly allocated to a training set and a validation set at a 7:3 ratio to develop an SVM-based model for estimating the 5-year overall survival (OS) in muscle-invasive bladder cancer. The model demonstrated robust predictive performance, with an area under the ROC curve (AUC) of 0.861, an accuracy of 0.889, a precision of 0.790, and an F1 score of 0.837. Decision curve analysis (DCA) confirmed the model’s significant net clinical benefit. Through integrated analyses of TCGA (n = 406) and GEO (n = 175) datasets, we identified 500 DEGs, which were refined to 240 radiomics-associated DEGs based on SVM-derived risk stratification. Immune infiltration analysis revealed that levels of NK cells, M0 macrophages, memory B cells, and activated dendritic cells were significantly higher in normal tissues compared to tumor tissues, indicating an immunologically favorable tumor microenvironment. Additionally, pathway analysis showed significant enrichment of immune-related pathways among the radiomics-associated genes.

conclusionWe identified gene markers associated with radiomic features, which may yield biological insights into imaging phenotypes and could potentially support the prognostic relevance of radiomics-based models.