A Deep Learning Approach for Consciousness Level Assessment in Sleep Stages: An Case framework Study

1479 

Abstract Submission 

Yong Liu¹, Gan Huang², Jinfei Tian³, Yao Wang⁴, Kin Cheung Lee⁵, Junling GAO⁶

¹Department of Critical Care Medicine, Shenzhen Hospital of Southern Medical, shenzhen, shenzhen, ²Shen Zhen University, Shenzhen, AK, ³Department of Critical Care Medicine, Shenzhen Hospital of Southern Medical, Shenzhen, AR, ⁴The University of Hong Kong, Hong Kong, AZ, ⁵The University of Hong Kong, Hong Kong, Hong Kong, ⁶The University of Hong Kong, Hong Kong, AS

Yong Liu  
Department of Critical Care Medicine, Shenzhen Hospital of Southern Medical
shenzhen, shenzhen

Gan Huang  
Shen Zhen University
Shenzhen, AK

Jinfei Tian  
Department of Critical Care Medicine, Shenzhen Hospital of Southern Medical
Shenzhen, AR

Yao Wang  
The University of Hong Kong
Hong Kong, AZ

Kin Cheung Lee  
The University of Hong Kong
Hong Kong, Hong Kong

Junling GAO  
The University of Hong Kong
Hong Kong, AS

Consciousness level assessment in clinical settings often relies on the Glasgow Coma Scale, which tends to be arbitrary. Additional information from EEG data could provide more objectivity. With this in mind, Lee et al. developed an explainable consciousness index (ECI) using deep learning methods. This study applies the ECI and the associated deep learning algorithm to sleep EEG data to assess its ability to categorize different sleep stages, each assumed to represent different levels of consciousness.

The ECI method, employing deep learning and the Layer-Wise Relevance Propagation (LRP) toolbox, was used to categorize overnight sleep EEG data. The C4 channel was selected, as previous studies have indicated that parietal EEG is likely related to consciousness activities. Sleep stages were labeled by clinicians for each 30-second EEG segment. The EEG data were first preprocessed by filtering at 0.1-45 Hz. Corrupted segments were replaced with adjacent EEG data, as independent component analysis could not be applied to single-channel EEG data or those with limited channels.

The data processing procedure was as follows:


Consciousness levels for the stages of wakefulness, rapid-eye-movement, N1, N2, and N3 were arbitrarily assigned values of 1, 0.6, 0.4, 0.3, and 0.2, respectively. The adapted ECI deep learning algorithm was then used to calculate the consciousness level for each 30-second stage. A correlation analysis was performed on the results from the clinical labels and the ECI-calculated results, yielding a correlation coefficient (r) of -0.199. When corrupted EEG segments were not replaced with adjacent EEG data, the correlation dropped to -0.132.

The application of the ECI algorithm to single-channel sleep data was not as effective as desired. This could be due to the lack of spatial information in single or limited-channel EEG data from ordinary sleep, which may weaken the ECI algorithm's ability to categorize consciousness levels. Additionally, the consciousness levels in sleep stages may not align directly with those seen in clinical conditions, such as the comparison between wakefulness and coma. Clinicians can often observe sleep and wakefulness even in coma patients. Corrupted EEG data may contain movement information, and additional information from heart rate, breathing rate, and body movement could be analyzed by the ECI algorithm, potentially improving its applicability to sleep data.

Classification and Predictive Modeling ¹

Consciousness and Awareness ²

Sleep and Wakefulness

Consciousness

Electroencephaolography (EEG)

Machine Learning