5 A multivariate plasma biomarker model enabling safe exclusion of glioblastoma
TL;DR
A plasma-based multivariate biomarker model enables accurate, safe exclusion of glioblastoma and supports diagnostic triage alongside neuro-imaging and maintains an extremely high NPV (>0.999), consistent with its intended role as a high-sensitivity exclusion tool rather than a confirmatory test.
Abstract
Glioblastoma (GBM) is the most aggressive primary brain tumor and remains difficult to diagnose due to reliance on MRI and invasive biopsies, which are limited by accessibility, morbidity, and radiologic uncertainty. Blood-based biomarkers may provide a minimally-invasive strategy to support early diagnostic-triage. In this retrospective case-control study, plasma samples from 72 patients with newly-diagnosed GBM and 102 age-matched healthy controls were analyzed. Seven circulating proteins were quantified using ultrasensitive- and clinical-grade immunoassays: glial fibrillary acidic protein (GFAP), neurofilament light chain (NfL), tau, neuron-specific enolase (NSE), myeloperoxidase (MPO), and polymorphonuclear-(PMN) elastase. Diagnostic performance was assessed using receiver-operating-characteristic (ROC) analysis. Multivariate logistic regression with 10-fold cross-validation and 1,000-iteration bootstrap resampling evaluated model robustness, with emphasis on sensitivity and negative predictive value (NPV) for exclusion-use. Among individual plasma biomarkers, GFAP demonstrated near-perfect separation between GBM and controls (AUC = 0.995), followed by NfL (AUC = 0.957), whereas tau, NSE, MPO, and PMN-elastase showed more modest individual performance. A multivariate logistic regression model integrating GFAP, NfL, tau, NSE, MPO, and PMN-elastase achieved near-perfect classification of GBM-versus-controls (AUC = 0.999), outperforming all individual biomarkers. At the optimal probability threshold, sensitivity reached 98.6% with 100% specificity, with no GBM cases misclassified as negative. Ten-fold cross-validation yielded a mean AUC of 0.984 with consistently high-sensitivity and specificity across folds, indicating minimal overfitting. Bootstrap resampling (1,000-iterations) confirmed stability of discriminative performance. When extrapolated to population-level prevalence, the model maintained an extremely high NPV (>0.999), while PPV remained low, consistent with its intended role as a high-sensitivity exclusion tool rather than a confirmatory test. A plasma-based multivariate biomarker model enables accurate, safe exclusion of glioblastoma and supports diagnostic triage alongside neuro-imaging.
