Comparison of logistic regression, support vector machines, and deep learning classifiers for predicting memory...

Area of Science:

• Neuroscience
• Machine Learning
• Biomedical Engineering

Background:

• Episodic memory encoding is crucial for learning and cognition.
• Predicting memory success from brain activity can inform memory enhancement technologies.
• Previous studies have explored machine learning for brain-computer interfaces.

Purpose of the Study:

• To compare logistic regression, support vector machines, and deep learning for predicting memory encoding success.
• To evaluate the impact of dimensionality reduction (tSNE) and feature selection on classifier performance.
• To inform the development of closed-loop brain stimulation devices for memory improvement.

Main Methods:

• Utilized stereo EEG data from 30 epilepsy patients performing a memory task.
• Trained and compared three binary classification models: logistic regression, SVM, and deep learning.
• Assessed the effects of feature reduction, including brain region selection and tSNE.

Main Results:

• Deep learning classifiers demonstrated superior performance compared to SVM and logistic regression.
• Selecting core brain regions improved logistic regression and SVM performance, especially with tSNE.
• Classifier performance was evaluated using AUC at both individual and aggregate levels.

Conclusions:

• Deep learning offers a powerful approach for predicting memory encoding success from EEG data.
• Feature engineering, including dimensionality reduction and targeted brain region selection, can optimize memory prediction models.
• These findings have implications for designing advanced brain-machine interfaces to modulate memory.