Towards stimulation-free automatic electrocorticographic speech mapping in neurosurgery patients

Area of Science:

• Neuroscience
• Machine Learning
• Neurosurgery

Background:

• Precise mapping of speech functions is vital for neurosurgical interventions in epilepsy and brain tumor patients.
• Traditional electrocortical stimulation mapping (ESM) is effective but carries a risk of inducing seizures.

Purpose of the Study:

• To develop and validate machine learning (ML) approaches for converting electrocorticographic mapping (ECM) signals into reliable speech function maps.
• To provide a safer alternative to invasive ESM for eloquent cortex identification.

Main Methods:

• A comprehensive ECM + electrocortical stimulation mapping (ESM) dataset was created from 14 patients undergoing stereo-EEG monitoring.
• Various ML classifiers (XGBoost, SVC, logistic regression, random forest, etc.) were trained using gamma-band power features from ECM signals.
• Model performance was validated using a leave-one-patient-out procedure, evaluating ROC and Precision-Recall curves.

Main Results:

• Linear support vector classification (SVC) achieved high accuracy (ROC-AUC 0.91, PR-AUC 0.88) in distinguishing speech-related iEEG channels.
• Incorporating voice onset information significantly improved classification accuracy.
• The study rigorously compared ECM and ESM results, mimicking real-world application.

Conclusions:

• Machine learning models can effectively predict speech function from ECM data, offering a safer alternative to ESM.
• The developed methods and publicly accessible dataset advance the field of eloquent cortex mapping.
• This research paves the way for safer and more reliable neurosurgical procedures.