Machine Learning Versus Logistic Regression Methods for 2-Year Mortality Prognostication in a Small, Heterogeneous...

Area of Science:

• Neuro-oncology
• Computational biology
• Biostatistics

Background:

• Machine learning (ML) traditionally requires large datasets but is increasingly applied to smaller, complex medical databases.
• Gliomas, a type of brain tumor, are graded using histopathology and now molecular characteristics.
• Accurate prognostication is crucial for patient management and treatment planning.

Purpose of the Study:

• To investigate the efficacy of ML techniques for predicting 2-year mortality in glioma patients.
• To assess ML performance on a small, high-dimensional dataset.
• To compare ML models with traditional logistic regression.

Main Methods:

• Applied artificial neural networks (ANNs), decision trees (DTs), and support vector machines (SVMs) to a glioma dataset of 76 patients.
• Compared model performance using raw data versus data with selected statistically significant features.
• Utilized logistic regression (LR) as a benchmark for comparison.

Main Results:

• ML models achieved reasonable predictive performance, with ANNs showing the highest accuracy (73.4%) after feature selection.
• Feature selection, reducing variables from 21 to 14, generally improved or maintained model performance across ML techniques.
• Performance metrics, including accuracy and area under the curve, were comparable across different ML algorithms and logistic regression.

Conclusions:

• Machine learning techniques are applicable and effective for prognostication in small, high-dimensional glioma datasets.
• Feature selection can enhance the performance of ML models in this context.
• ML offers a valuable tool for analyzing limited medical data, complementing traditional statistical methods.