Non-tumor plasma extracellular vesicles predict the occurrence of gbm: a machine learning analysis

Abstract INTRODUCTION Glioblastoma (GBM) is the most common primary malignant brain tumor and has a poor clinical outcome. Extracellular vesicles (EVs) are promising biomarkers to assist the differential diagnosis between pseudoprogression and true tumor progression. We aimed to identify non-tumor plasma EV markers that predicted GBM occurrence using trained multiple tree-based machine learning algorithms. METHODS We included plasma samples of 40 GBM and 40 non-GBM (20 brain metastasis, 20 normal donors). Spectral flow cytometry was performed to determine plasma EV subpopulation frequency based on CD9,CD81,CD63,CD11b,CD45,CD31, and CD41a expression. Multiparametric t-SNE and FlowSOM analysis clustered plasma EVs based on these markers. Decision trees and random forest classifiers were trained and tested to predict the disease status (GBM vs non-GBM) based on plasma EV subpopulations. RESULTS Compared to non-GBM, GBM had increased frequency of CD9+ EVs (64.52% vs 40.69%, p< 0.0001), CD9+CD63-CD81- EVs (54.45% vs 37.26%, p< 0.0001) and myeloid-derived EVs (CD9+CD11b+ EVs, 20.77% vs 12.48%, p = 0.0241). In the decision tree analysis, low frequency of CD9+ EVs was found in 44% of all cases and represented a chance of 75% to be classified as non-GBM cases. An algorithm based on random forest analysis (n = 500 trees) using these markers had a sensitivity of 85.71% and specificity of 77.8% with an overall accuracy of 81.25% for identifying GBM. The frequencies of CD9+, CD9+CD63-CD81-, and CD9+CD63+CD81+ were the most critical factors in predicting GBM diagnosis. FlowSOM analysis identified 20 plasma EV subpopulations but had a lower accuracy (62.5%) when combined with our machine learning algorithm. CONCLUSIONS Non-tumor plasma EV expression profile combined with tree-based machine learning demonstrated high accuracy in predicting patient disease status. CD9+ EV frequency was one of the most promising biomarkers to assist in this diagnosis. We will continue to test our established algorithm and further validate our findings in a larger cohort.