Optimized Multilayer Perceptron for Sensorimotor Functional Mapping Based on a Few Minutes of Intracranial...

Area of Science:

• Computational Neuroscience
• Epilepsy Research
• Machine Learning in Medicine

Background:

• Accurate localization of the sensorimotor cortex (SM) and central sulcus (CS) is crucial for surgical planning in intractable epilepsy.
• Current methods may have limitations in precision and invasiveness.
• Intracranial electroencephalogram (icEEG) offers high-resolution brain activity data during sleep.

Purpose of the Study:

• To develop and evaluate an optimized multilayer perceptron (MLP) neural network for mapping the SM and identifying the CS.
• To assess the MLP's performance using multiple quantitative metrics.
• To investigate the impact of feature extension and weighted imbalanced data on MLP efficacy.

Main Methods:

• Utilized 6-minute free-running icEEG recordings during sleep in patients with intractable epilepsy.
• Developed an MLP neural network with specific layer configurations (4 neurons, hyperbolic TanH and Gaussian activation functions).
• Employed 10-fold cross-validation and evaluated performance using accuracy, AUC, recall, precision, F1-scores, and specificity.

Main Results:

• The optimized MLP model demonstrated accurate mapping of the SM and identification of the CS.
• Feature extension (ε) and weighted imbalanced data (w) significantly improved MLP performance metrics.
• The study provides a robust evaluation of the MLP's efficacy in a clinical neurophysiological context.

Conclusions:

• An optimized MLP neural network is a viable tool for precise neuroanatomical mapping in epilepsy patients using icEEG.
• The proposed method offers potential for improved pre-surgical evaluation in intractable epilepsy.
• Further research can explore broader applications of this machine learning approach in neurosurgery.