Prognostic Modeling for Brain Cancer Patient Survival using Advanced Supervised Machine Learning Classification Approaches

Creative Commons License

Dirican O.

VOPROSY ONKOLOGII, cilt.7, sa.2, 2025 (Scopus)

Özet

Aim: Survival prediction in brain cancer is crucial for guiding treatment decisions and improving patient outcomes. This study aimed to develop a novel prognostic model for brain cancer patient survival using advanced supervised machine learning classification approaches. The model incorporated demographic, clinical, immunohistochemical, and genomic characteristics.

Methods: A retrospective examination was conducted on a cohort of 149 patients with intracranial tumors who underwent operative procedures. Systematic collection of demographic and clinical data was performed. Tumor and adjacent healthy tissues underwent histopathological scrutiny and immunohistochemical analysis for GST-P, GST-T, GST-M, CYP1A1, CYP1B1, MDR, and p53 expression. Staining intensity was graded, and genomic DNA from tumor tissues was analyzed for GSTM1, GSTT1, and p53 genotypes. Five categories of input features, comprising demographic, clinical, protein expression profiles, and genotypic information, were selected for model development. Totally 15 models were constructed via decision tree, naïve bayes and SVM, using Python, and compared for accuracy, precision, sensitivity, and F-measure metrics.

Results: The overall postoperative survival rate was 65%. Protein expression analysis revealed significant differences between cancerous and normal tissues for GST-P, GST-T, GST-M, CYP1A1, CYP1B1, MDR, and p53 genes. Genotyping indicated that the GST-M1 null genotype might be associated with brain tumor development. Model evaluation demonstrated a decision tree accuracy of 84%, the highest among models incorporating demographic, clinical, immunohistochemical, and genetic information. Precision and sensitivity varied, but the decision tree showed acceptable values for accuracy, precision, and sensitivity.

Conclusions: Decision trees may be suitable for predicting brain cancer survival particularly in limited datasets, compared to SVM and Naive Bayes when using demographic, clinical, immunohistochemical, and genotypic variables.