Gliomas exhibit significant prognostic heterogeneity, and single-omics data/existing technologies struggle to balance multi-omics integration efficiency, prediction accuracy, and clinical adaptability-hindering the clinical translation of precise prognostic assessment. Focussing on glioblastoma (GBM) and lower-grade glioma (LGG), this study proposes an integrated solution: three-step multi-omics feature selection combined with the limited random forest (FRF) model, using TCGA-derived transcriptomic, genomic, epigenomic and clinical survival data. Prognosis-related features are first screened using univariate Cox regression, refined by random forest-based feature importance for dimensionality reduction and then integrated into a multi-omics matrix through sample matching. The FRF model balances efficiency and accuracy by limiting decision tree number and depth, optimising node splitting criteria, and adding a dual-weighted correction mechanism. Results show that FRF achieves an AUC of 0.96 for GBM, outperforms logistic regression (LR) and support vector machine (SVM) across all LGG metrics, and reduces training time to minutes-meeting the 2-h clinical prognostic demand. Ablation experiments confirm that the multi-omics model improves performance by 11.63% compared with the optimal single-omics model, with core features consistent with glioma molecular mechanisms. This resolves the challenge of rapid integration and precise prediction, providing an efficient tool for glioma prognostic assessment and supporting multi-omics clinical translation.
Liu et al. (Thu,) studied this question.