Abstract Background While both MRI radiomics and miRNA liquid biopsy have shown promise for glioblastoma prognostication, state-of-the-art methods may enable novel integration of multimodal data to further optimize performance. Methods Serum samples (n = 193) were collected from 73 patients with pathology-confirmed glioblastoma at a single center. We quantified 798 miRNAs per sample using nCounter. Radiomic features were automatically extracted from MRI scans (n = 306) obtained during the same follow-up period. A new data integration pipeline was applied to evaluate machine learning-based outcome predictions using miRNA, radiomic, and miRNA+radiomic input datasets. All models included a set of clinical covariates of known prognostic value. Time-averaged AUC analysis was used to incorporate longitudinal sampling. In experiments to classify postoperative samples labeled by likely disease burden, performance-driving miRNAs were further investigated using counterfactual analysis (CA) and Shapley Additive Explanations (SHAP). Results The recurrence rate was 75% (median 222 days, IQR 93–378), and post-recurrence mortality was 62% (412 days, 274–695) during the follow-up period. Radiomics outperformed miRNAs for recurrence prediction (time-averaged AUC = 0.66 0.60–0.71 versus AUC = 0.56 0.53–0.59), while miRNAs outperformed radiomics for survival prediction (AUC = 0.70 0.64–0.77 versus AUC = 0.55 0.47–0.60). Combined miRNA+radiomics models performed well for recurrence (AUC = 0.64 0.44–0.80) and best overall for survival (AUC = 0.76 0.63–0.86). Several performance-driving miRNAs were also identified on CA and SHAP analyses. Conclusions Serum miRNA profiling combined with MRI radiomics may improve longitudinal approximation of postoperative glioblastoma prognosis. Integrating multimodal data is feasible and could enable more informed counseling of patients and families over the disease course.
Leary et al. (Thu,) studied this question.