A comparison of machine learning classifiers for pediatric epilepsy using resting-state functional MRI latency data.

Area of Science:

• Neuroscience
• Medical Imaging
• Artificial Intelligence

Background:

• Epilepsy is a common pediatric neurological condition impacting brain development.
• Early and accurate diagnosis is crucial for effective seizure control and management.
• Resting-state functional MRI (rfMRI) offers insights into brain network function.

Purpose of the Study:

• To evaluate machine learning algorithms for epilepsy detection using rfMRI latency data.
• To compare the performance of XGBoost, SVM, and Random Forest models.
• To assess the potential of these AI tools as adjunctive diagnostic methods.

Main Methods:

• Acquired preoperative rfMRI and anatomical MRI scans from 63 epilepsy patients and 259 controls.
• Analyzed latency z-score distributions and overlap in 36 seed regions.
• Extracted features using principal component analysis and trained XGBoost, SVM, and Random Forest models.

Main Results:

• The XGBoost model achieved the highest performance with an AUC of 0.79, 74% accuracy, 73% specificity, and 77% sensitivity.
• Random Forest showed comparable performance in some metrics but lower sensitivity (31%).
• SVM performance did not exceed 70% in key metrics.

Conclusions:

• Machine learning algorithms trained with rfMRI latency data can effectively aid in epilepsy detection.
• The XGBoost model shows significant potential for improving epilepsy diagnosis accuracy and sensitivity.
• This approach could facilitate timely and appropriate care for pediatric epilepsy patients.