Multiple Time Series Fusion Based on LSTM: An Application to CAP A Phase Classification Using EEG.

Area of Science:

• Neuroscience
• Signal Processing
• Machine Learning

Background:

• The Cyclic Alternating Pattern (CAP) is a key indicator of sleep instability detected in EEG signals.
• Clinical applications of CAP assessment are hindered by the lack of automated detection methods.
• Developing automatic methodologies is crucial for real-world CAP analysis.

Purpose of the Study:

• To propose and evaluate a deep learning-based approach for automatic CAP A phase classification.
• To optimize channel selection, feature fusion, and classification using two distinct algorithms.
• To assess the performance of the developed methodology on a challenging dataset.

Main Methods:

• A deep learning model was employed for feature-level fusion of EEG channels.
• Two optimization algorithms were utilized to refine channel selection and fusion strategies.
• The methodology was validated using EEG data from patients with and without nocturnal frontal lobe epilepsy.

Main Results:

• Both optimization algorithms identified a consistent set of three EEG channels (Fp2-F4, C4-A1, F4-C4) for optimal CAP detection.
• The optimized models achieved an Area Under the Receiver Operating Characteristic Curve (AUC) of 0.82 and an average accuracy of 77-79%.
• The models demonstrated robustness against noise and channel loss, performing comparably to specialist agreement.

Conclusions:

• The proposed deep learning methodology enables fully automatic CAP A phase classification.
• The system is resilient to noise and channel loss, making it suitable for real-world applications.
• This automated approach facilitates broader clinical use of CAP assessment for sleep instability.