Explainable AI for Intraoperative Motor-Evoked Potential Muscle Classification in Neurosurgery: Bicentric...

Area of Science:

• Neurosurgery
• Neurophysiology
• Artificial Intelligence

Background:

• Intraoperative neurophysiological monitoring (IONM) is crucial for preserving motor pathways during high-risk surgeries.
• Current methods lack standardized warning criteria for motor-evoked potentials (MEPs), impacting patient safety.
• Developing muscle identification prediction models can enhance IONM reliability.

Purpose of the Study:

• To develop and evaluate machine learning (ML) models for muscle classification in a bicentric setup.
• To identify key MEP signal features for accurate muscle classification using explainable artificial intelligence (XAI).

Main Methods:

• Random forest (RF) classifiers and convolutional neural networks (CNNs) were used to classify MEP signals from four muscles.
• Models were trained and validated on data from 151 surgeries and tested on data from 58 surgeries across two centers.
• Shapley Additive Explanation (SHAP) and gradient class activation maps (Grad-CAM) were employed for feature identification.

Main Results:

• The RF classifier achieved 87.9% accuracy on the validation set and 80% on the test set.
• 1D- and 2D-CNNs showed comparable performance in MEP classification.
• XAI analysis revealed frequency components and peak latencies as critical features for accurate classification.

Conclusions:

• ML techniques and XAI significantly enhance the reliability of AI-driven IONM.
• This study identified novel MEP signal features that could improve intraoperative warning criteria.
• Reduced muscle mislabeling and new warning criteria can enhance patient safety in surgical procedures.