Machine-Learning-Based Depression Detection Model from Electroencephalograph (EEG) Data Obtained by Consumer-Grade...

Area of Science:

• Neuroscience
• Machine Learning
• Computational Psychiatry

Background:

• Machine learning applied to electroencephalograph (EEG) data shows promise for depression detection.
• Medical-grade EEG has demonstrated accurate depression detection capabilities.
• A gap exists in achieving comparable accuracy with simpler, consumer-grade EEG sensors.

Purpose of the Study:

• To enhance depression detection accuracy using machine learning with consumer-grade EEG sensors.
• To identify and select optimal EEG indices for improved depression detection.
• To validate the efficacy of machine learning models trained on consumer-grade EEG data.

Main Methods:

• Quantified EEG indices including power spectrum, asymmetry, complexity, and functional connectivity.
• Employed feature selection methods (LightGBM, mutual information, ReliefF, ElasticNet) to identify key EEG indices.
• Trained a Light Gradient Boosting Machine (LightGBM) model on selected EEG indices, ensuring data independence through cross-validation.

Main Results:

• Achieved a Macro F1 score of 91.59% for depression detection using consumer-grade EEG.
• Identified specific EEG indices, like differential entropy and functional connectivity, yielding approximately 80% Macro F1 score individually.
• Demonstrated the potential of consumer-grade EEG data in machine learning-based depression detection.

Conclusions:

• Consumer-grade EEG sensors, when analyzed with machine learning, can effectively detect depression.
• Selected EEG indices show significant promise for future depression detection applications.
• This approach offers a more accessible pathway for mental health monitoring using brainwave technology.