Machine Learning Predicts Short-term vs Long-term Survival Using Biomarker Profile in Glioblastoma

Introduction: Glioblastoma multiforme (GBM) is the most common primary, malignant brain tumor and continues to have a grim prognosis of roughly 7.2% of patients surviving to five years after diagnosis. Prognosis prediction in GBM remains challenging due to the complex interplay of demographic and molecular factors. Biomarkers have emerged as essential tools in the prognostication of GBM, offering potential to enhance predictive accuracy and personalize treatment strategies. This study aims to evaluate machine learning (ML) approaches for predicting GBM patient survival, utilizing demographic data and biomarker profiles to classify short- and long-term survival outcomes.

Methods: This is an IRB-approved retrospective chart review of patients treated at our institution, from 2017 to the present. Demographic information (age, race, gender), surgical resection, biomarker profiles from neuropathology reports, and survival outcomes have been and are currently being collected. Data is being analyzed with R.

Results: A stochastic gradient descent regressor was able to predict progression-free survival with R^2 of 0.7566, with minimal hyperparameter tuning. Further optimization may improve results, and other regression techniques can also be used to optimize performance.

Conclusion : Regression models are able to correlate biomarkers with progression-free survival with moderate degree of accuracy. This is our preliminary report and further analyses are in progress. A machine learning tool can help clinicians prognosticate GBM more accurately in the future.