Abstract
This paper presents a fusion-based neural network (NN) classification algorithm for 40-Hz auditory steady state response (ASSR) ensemble averaged signals which were recorded from eight human subjects for observing sleep patterns (wakefulness $W_0$ and deep sleep $N_3$ or slow wave sleep $SWS$). In $SWS$, sensitivity to pain is the lowest relative to other sleep stages and arousal needs stronger stimuli. 40-Hz ASSR signals were extracted by averaging over 900 sweeps on a 30-s window. Signals generated during $N_3$ deep sleep state show similarities to those produced when general anesthesia is given to patients during clinical surgery. Our experimental results show that the automatic classification system used identifies sleep states with an accuracy rate of 100% when the training and test signals come from the same subjects while its accuracy is reduced to 97.6%, on average, when signals are used from different training and test subjects. Our results may lead to future classification of consciousness and wakefulness of patients with 40-Hz ASSR for observing the depth and effects of general anesthesia (DGA).
Introduction
Monitoring sleep stages and levels of consciousness is critical in both sleep medicine and clinical anesthesiology. Traditionally, this relies on manual scoring of Electroencephalogram (EEG) data, a process that is time-consuming and prone to human error. While automated EEG classification exists, the high dimensionality and inherent noise of raw brainwave activity make it difficult for algorithms to generalize across different individuals.
This research explores a robust alternative: the 40-Hz Auditory Steady State Response (ASSR). Unlike broad EEG activity, the ASSR is a specific electrical signal the brain produces in response to rapid auditory stimuli. Because its amplitude is highly sensitive to arousal levels—dropping significantly during deep sleep ($N_3$) and general anesthesia—it serves as an excellent proxy for monitoring consciousness.
Methodology
Signal Extraction and Fusion
The researchers recorded EEG data from 11 scalp sites on eight human subjects using the standard 10-20 system. To isolate the 40-Hz ASSR signal from background neural noise, they employed weighted ensemble averaging. This involved averaging 900 "sweeps" (short segments of data) over a 30-second window, using Kalman filter coefficients to weight the data for maximum clarity.
A key innovation in this study is Feature-Level Fusion. Rather than relying on a single electrode, the system concatenates data from two separate channels (e.g., frontal and central sites). This redundancy ensures that if one channel is compromised by noise or artifact, the other can maintain the integrity of the classification.
Classification Models
Two primary machine learning models were compared:
- Learning Vector Quantization (LVQ): An artificial neural network that uses "prototype" vectors to represent classes.
- Support Vector Machines (SVM): A standard supervised learning model used for binary classification.
Key Results
The models were evaluated using two protocols: intra-subject (training and testing on the same individual) and inter-subject (training on a group and testing on a new individual).
- Perfect Intra-subject Accuracy: The LVQ system achieved 100% accuracy when identifying wakefulness vs. deep sleep for the same subject.
- High Generalization: In inter-subject testing, the LVQ system maintained an average accuracy of 97.6%, demonstrating its ability to handle biological variability.
- LVQ vs. SVM: LVQ significantly outperformed SVM in cross-subject tests. While LVQ remained stable, SVM error rates often exceeded 20%, proving it less capable of defining generalized "prototypes" for sleep states across different brain anatomies.
Clinical Implications and Future Work
The biological similarity between $N_3$ deep sleep and the state induced by general anesthesia suggests that this 40-Hz ASSR classifier could be adapted into a Depth of General Anesthesia (DGA) monitor. Such a tool would provide clinicians with a reliable, automated measure of patient consciousness during surgery.
Future research will aim to expand the system's capabilities to include other sleep stages (REM, $N_1$, $N_2$) and increase the subject pool to further validate these findings for broad clinical application. Additionally, reducing the 30-second processing window will be essential for real-time monitoring in surgical environments.
Summary updated