Gliomas are the most common primary tumor of the central nervous system, and glioblastomas are the IV. and most aggresive stage. The purpose of this master's thesis is to analyze the factors that influence the survival of patients diagnosed with glioblastoma, based on a given data set collected by the Trans-Glioma project. The key focus of this task is primarily interpretability since we are more interested in the connections between survival itself and the predictive variables rather than the actual use of these models in the future.
We begin the thesis with a theoretical introduction about gliomas and specifically about glioblastomas. We describe the historical development of classification methods and explain the biological background, including the role of genetic predispositions. We also check the influence of the expression of certain genes on survival.
We continue with a theoretical background of statistical and machine learning methods which set the framework for subsequent data analysis. In the thesis we look into various statistical and machine learning techniques such as logistic regression, naive Bayes, random forests, and XGBoost highlighting their strengths and limitations in the context of one-year survival classification with a glioblastoma diagnosis.
After the theoretical presentation we continue with the presentation of the collected data, accompanied by statistical analysis and auxiliary visual displays.
We explain the statistical modeling procedure, first presenting the set of independent variables and then describing the methodology itself used to develop predictive models for classifying one-year survival at diagnosis of glioblastoma. This includes data preprocessing, model preparation, and the implementation method nested within this task is 10-fold cross-validation.
We conclude with an analysis of the results by comparing accuracy metrics such as precision, recall, precision, and F1-value between the models, as well as the distribution of the importance of the predictor variables between the models. We find that the patient's age is a very important indicator of survival, and we also claim this for the epidermal growth factor receptor (EGFR) mutation. In the case of various genes (ASCBG1, CCR5, CD9, STAT3, etc.), we check whether their expression affects survival, and we find that expression has a significant impact on predictions. We highlight the expression of the CD9, S100A4, STAT3 and STMN4 genes as the most important.
The findings of this research contribute to a deeper understanding of glioblastomas and the potential use of statistical and machine learning techniques in the classification of one-year or any survival.
|
