Heuristic Optimization of Deep and Shallow Classifiers: An Application for Electroencephalogram Cyclic Alternating...

Area of Science:

• Neuroscience
• Sleep Medicine
• Computational Biology

Background:

• Accurate analysis of sleep stages, including non-rapid eye movement (NREM) sleep and cyclic alternating patterns (CAPs), is crucial for diagnosing sleep disorders.
• Current methods for CAP analysis can be labor-intensive and subjective.
• Developing automated, objective methodologies is essential for large-scale sleep research and clinical applications.

Purpose of the Study:

• To propose and evaluate automated methodologies for analyzing electroencephalogram (EEG) signals for non-rapid eye movement (NREM) sleep and cyclic alternating pattern (CAP) assessments.
• To compare the performance of different machine learning models, including convolutional neural networks (CNNs), long short-term memory (LSTM) networks, and feed-forward neural networks (FFNNs), for sleep stage classification.
• To develop and apply hyper-parameter tuning algorithms to optimize classifier performance.

Main Methods:

• EEG signals from a single monopolar derivation were analyzed.
• Machine learning models (1D-CNN, LSTM, FFNN) were trained for NREM and A-phase classifications.
• A finite state machine was used for CAP cycle scoring.
• Two hyper-parameter tuning algorithms were developed to optimize the classifiers.
• The study included both healthy subjects and those with sleep-disordered breathing.

Main Results:

• The LSTM model fed with proposed features demonstrated superior performance.
• For A-phase classification, the LSTM model achieved 83% accuracy and an 0.88 area under the receiver operating characteristic curve (AUC).
• For NREM estimation, the LSTM model achieved 88% accuracy and an 0.95 AUC.
• The LSTM model achieved 79% accuracy for CAP cycle classification, with a 22% percentage error for CAP rate.

Conclusions:

• Automated analysis of NREM sleep and CAPs using LSTM networks is feasible and effective.
• The proposed methodologies offer a promising approach for objective and efficient sleep analysis.
• Further research can refine these models for improved clinical diagnostic capabilities in sleep medicine.