Bewertung des Potentials eines Machine Learning Models mit MGMT-Methylierung als Hauptvariable bei Glioblastoma-Patienten in einem limitierten Datenset

Machine learning is increasingly finding applications in glioblastoma (GBM) research, focusing on large datasets derived from genetic analysis and imaging for predictive modeling.

This study shifts attention to fundamental patient data, particularly exploring the impact of MGMT-Methylation on GBM patient prognosis. With this, we aim to assess the predictability of overall survival, using a machine learning framework.

The patient cohort comprised retrospective data from 218 patients from a single-center study. Age, sex and MGMT methylation were used to predict overall survival of patients in months. The predictive performance was measured in mean absolute error (MAE), mean squared error (MSE), explained variance (EV), Pearson correlation and r-squared (r2) and evaluated within a 10x10 nested cross-validation framework. The machine learning pipeline comprised five regression estimators, including both linear and non-linear approaches. The variables were further evaluated based on their impact on the prognosis using feature importance analysis. Statistical significance was assessed using a permutation test procedure.

The machine learning pipeline achieved an MAE of 12.88 (SD=2.18) and an EV of 3% (SD=1.8%) when predicting survival duration in months. It marginally outperformed a baseline model that always predicts the mean survival of the training sample (MAE=13.66 months). Our analysis showed that the linear method of Support Vector Regression led to more accurate predictions compared to non-linear estimators. Feature importance testing indicated that age (permutation norm=0.897) and positive MGMT-Methylation (permutation norm=0.045) have the greatest influence on the model predictions.

Our findings emphasize the efficacy of a linear machine learning approach on a limited dataset, with focus on MGMT-Methylation's impact on the prognosis of glioblastoma patients in terms of overall survival rated in months.

This method shows potential especially in accurately predicting long-term survival based on the tumor's genetic profile. Such predictive capabilities could significantly assist surgeons and oncologists in their routine decision-making.

Fig.1: Support Vector Regression performance - Light blue: training dataset performance; Dark blue: Validation on test model

Fig.2: Predictions in training and test set - True value on y-axis, predicted overall survival in months on x-axis. Left: Training set (90% data). Right: Test set (10% data)