Predicting seizure outcome after epilepsy surgery: Do we need more complex models, larger samples, or better data?

Area of Science:

• Neurology
• Machine Learning in Medicine
• Epilepsy Surgery Outcomes

Background:

• Accurate prediction of seizure freedom post-epilepsy surgery remains a significant clinical challenge.
• Existing models often struggle to generalize, highlighting the need for improved predictive strategies.

Purpose of the Study:

• To investigate if complex models, larger sample sizes, or data-driven feature selection enhance prediction of postoperative seizure outcome.
• To perform external validation of a machine learning model for predicting seizure freedom.

Main Methods:

• Retrospective cohort study of 797 pediatric epilepsy surgery patients.
• Trained and evaluated logistic regression, multilayer perceptron, and XGBoost models.
• Assessed impact of sample size via learning curves and evaluated data-driven feature selection.

Main Results:

• No significant performance difference between logistic regression, MLP, and XGBoost models (accuracy ~71-72%).
• All trained models outperformed an external XGBoost model (accuracy 63%).
• Performance improved with sample size, but plateaued; data-driven feature selection yielded the best results.

Conclusions:

• Neither complex machine learning models nor large datasets alone significantly improve seizure freedom prediction.
• Advancements require improved feature selection, data standardization, collaboration, and model sharing in the field.