BIOM-70. MIRNA PROFILING COMBINED WITH MRI RADIOMICS FOR GLIOBLASTOMA OUTCOME PREDICTION: INTEGRATED MACHINE LEARNING ANALYSIS OF LONGITUDINAL, MULTIMODAL DATA FROM 73 PATIENTS

Abstract INTRODUCTION miRNA liquid biopsy and MRI radiomics have both shown promise for prognostication in glioblastoma, though neither has been translated to clinical use. Leveraging a prospective, longitudinal, dataset from a major tertiary care center (2019–2024), we implemented a novel multimodal data integration pipeline and an unbiased approach to feature identification to evaluate the utility of miRNAs versus MR radiomics for predicting glioblastoma recurrence and survival. METHODS Serum samples were collected longitudinally from 73 patients with pathology-confirmed glioblastoma (193 total samples). 798 miRNAs were quantified per sample (count per 40 µL serum) using the Nanostring nCounter platform. Automated radiomic feature extraction was performed on MRI data obtained during the same follow-up period. A novel data integration pipeline was employed to evaluate miRNA, radiomic, and combined miRNA+radiomic data as input datasets to machine learning -based outcome prediction (80/20 test-train split). Multiple feature selection techniques were tested per input dataset. All models also included a set of demographic and clinical variables (age, gender, extent of resection, and MGMT promotor methylation). Time-averaged AUC analysis was utilized to incorporate longitudinal sampling in outcome prediction. miRNAs of importance were further confirmed using counterfactual analysis (CA) and SHapley Additive exPlanations (SHAP), which were applied to classify a subset of postoperative samples labelled by disease burden (early post-op versus likely recurrence) based upon miRNAs alone. RESULTS During the follow-up period, the recurrence rate was 75% (median 222 days, interquartile range [93–378]) and mortality was 62% (412 [274–695]). miRNA-only and radiomics-only models performed well for survival prediction (time-averaged AUC=0.77 [0.62–0.87] and 0.75 [0.56–0.93], respectively), but the fully integrated (miRNA+MRI) model offered the best performance (AUC=0.90 [0.81–0.99]). Results were similar for recurrence prediction. A subset of miRNAs (e.g., miR-320e, miR-451a) were observed as important across CA, SHAP, and recurrence prediction analyses. CONCLUSIONS Serum miRNA profiling may offer an inexpensive means of estimating glioblastoma prognosis, alone or in combination with modern radiomics-based methods.