Introduction: Glioma recurrence after surgical resection remains a significant challenge in neuro-oncology. Advances in liquid biopsy techniques offer a minimally invasive method for tracking tumor biomarkers. This study integrates liquid biopsy-derived biomarkers and pre-operative clinical data to develop machine learning models that predict glioma recurrence after surgical resection.
Methods: Extracellular vesicle (EV) based detection platforms were previously developed for detection of glioma specific TERTp (EV DNA + cfDNA), EGFRvIII (EV RNA), and IDH1.R132H (EV RNA) mutations. Results from the plasma testing of the three study cohorts were integrated to develop a predictive model. Different machine learning algorithms (Linear Model, LightGBM, XGBoost, an Artificial Neural Network, Random Forest, Decision Tree classifier, and ensemble model) were tested for their ability to predict glioma recurrence based on pre-operative metrics within one year from surgical resection (90:10 train:test split). Model performance was assessed using AUROC, accuracy, precision, recall, and F1 score.
Results: The mean glioma recurrence rate across the three cohorts (N=180 patients) within one year was 33.9% (N=61). Of the machine learning algorithms developed, the ensemble model (LightGBM and Neural Network combined) performed the best (AUC=0.949, F1=0.903, accuracy=0.923, precision=0.824, recall=1), followed by the Artificial Neural Network alone (AUC=0.943, F1=0.867, accuracy=0.897, precision=1, recall=1). Analysis of model behavior using SHAP feature importance uncovered that IDH1 mutation status, WHO Grade, and tumor location were the most important features.
Conclusion : Machine learning models based on solely pre-operative tumor characterstics and liquid biopsy data were able to predict postoperative glioma recurrence by one-year follow-up with high performance across evaluation metrics. This finding has important implications in outcome prognostication, and our machine learning models are promising tools towards precision oncology.
.jpg)
