A machine learning eye movement detection algorithm using electrooculography.

Area of Science:

• Neuroscience
• Sleep Medicine
• Artificial Intelligence

Background:

• Quantifying eye movements in polysomnograms (PSG) is challenging and time-consuming.
• Automated detection can facilitate studies on sleep-related eye movement patterns.
• Such studies may aid in diagnosing neurological disorders and monitoring treatment efficacy.

Purpose of the Study:

• To develop and validate a long short-term memory (LSTM) algorithm for automated eye movement detection in PSG.
• To assess the algorithm's sensitivity and specificity in identifying eye movement presence, direction, and speed.

Main Methods:

• A retrospective study utilized one-hour PSG samples from 47 patients.
• Manual identification of eye movements trained an LSTM algorithm.
• 5-fold cross-validation and a 'fuzzy' evaluation method were employed for robust assessment.

Main Results:

• Eye movements constituted 9.4% of the analyzed EOG recording time across all patients.
• The LSTM model achieved an average sensitivity of 0.88 and specificity of 0.89.
• The 'fuzzy' evaluation improved performance to 0.93 sensitivity and 0.92 specificity.

Conclusions:

• An automated algorithm effectively detects eye movements from EOG signals with high accuracy.
• Noninvasive, automated detection holds significant clinical potential for sleep stage classification.
• It can help establish normative eye movement distributions in various patient populations.