[Data mining in diagnostic knowledge acquisition from patients with brain glioma]

Area of Science:

• Neuroscience
• Medical Informatics
• Artificial Intelligence

Background:

• Brain glioma diagnosis requires accurate prediction of malignancy.
• Traditional diagnostic methods can be time-consuming and subjective.
• Data mining offers potential for automated and objective diagnostic support.

Purpose of the Study:

• To evaluate and compare the performance of three data mining algorithms for predicting brain glioma malignancy.
• To assess the accuracy, reliability, and interpretability of Multi-Layer Perceptron network (MLP), decision trees, and rule induction.
• To determine the feasibility of using data mining for computer-aided diagnosis systems in neuro-oncology.

Main Methods:

• Collected and preprocessed data from 280 brain glioma patient cases, handling missing values.
• Applied three data mining algorithms: Multi-Layer Perceptron network (MLP), decision tree, and rule induction.
• Conducted a 10-fold cross-validation test to compare algorithm performance based on accuracy, reliability, and interpretability.

Main Results:

• Multi-Layer Perceptron network (MLP) demonstrated the highest accuracy and reliability, particularly with few hidden nodes, though its results were less interpretable.
• Decision trees and rule induction provided more understandable and applicable diagnostic knowledge using attribute-value pairs.
• All algorithms achieved over 80% accuracy, meeting neuroradiologists' requirements, with rule induction favored for results requiring further evaluation.

Conclusions:

• Data mining techniques are effective in extracting valid diagnostic knowledge from brain glioma cases.
• MLP is optimal when diagnostic accuracy is paramount, while rule induction is preferred for interpretability and further clinical assessment.
• The study validates the potential for developing feasible computer-aided diagnosis systems for brain glioma.