Machine learning for (non-)epileptic tissue detection from the intraoperative electrocorticogram.

Area of Science:

• Neuroscience
• Medical Technology
• Artificial Intelligence in Medicine

Background:

• Clinical visual intraoperative electrocorticography (ioECoG) reading is crucial for localizing epileptic tissue and improving epilepsy surgery outcomes.
• The potential for machine learning (ML) to enhance ioECoG interpretation and identify key predictive features remains an area of active research.

Purpose of the Study:

• To investigate if ML can complement clinical ioECoG reading for epilepsy surgery.
• To determine how patient subgroups influence ML performance in ioECoG analysis.
• To identify the most significant ioECoG spectral features utilized by ML models.

Main Methods:

• Trained an extra trees classifier (ETC) using 14 spectral features on 71 patients (training set) with Engel 1A outcome post-surgery.
• Classified ioECoG channels as resected or non-resected tissue.
• Compared ETC performance against clinical ioECoG reading in a test set of 20 patients, utilizing explainable AI (xAI) to identify key features.

Main Results:

• The ETC showed comparable or superior performance to clinical reading in 14 out of 20 test patients.
• The ETC achieved the highest performance in the tumor subgroup (AUC: 0.84).
• Explainable AI identified relative theta, alpha, and fast ripple power as predictors of resected tissue, and relative beta and gamma power for non-resected tissue.

Conclusions:

• Subtle spectral ioECoG changes, not perceptible to the human eye, can assist in differentiating healthy from pathological tissue.
• ML models incorporating spectral ioECoG features can serve as a valuable adjunct to, rather than a replacement for, clinical ioECoG interpretation.
• The ML approach shows particular promise in supporting surgical decisions for patients with brain tumors.