Automatic Sleep Stage Classification using a Neural Network Algorithm Zoe Cohen Supervisor: Prateek Gundannavar, Principal Investigator: Arye Nehorai, PhD Department of Electrical and Systems Engineering Abstract: For this project I developed and tested a neural network algorithm for the purpose of performing automatic sleep stage classification. Sleep is typically classified into six different stages: wake, N1, N2, N3/N4, and REM (rapid eye movement). The classification is based on various standards set by the American Academy of Sleep Medicine (AASM) and requires a trained sleep technician. In this project I wrote a neural network algorithm to perform classification based on these standards, thus making the process automatic. The neural network algorithm was developed by improving and building on previous iterations, the final result being a classifier capable of discriminating between five different classes with 80.82% accuracy. Introduction: Feature Extraction: Results: Sleep scoring is the procedure by which various biological signals (EEG, EMG, EOG) are examined to determine the sleep stage of a patient. It is valuable for its diagnostic potential as abnormalities in sleep can serve as biomarkers for various diseases, such as Alzheimer’s. However, the procedure of sleep scoring requires the recording of complex biological signals and a trained sleep technician to review the recordings. A neural network algorithm would eliminate the need for a technician, thus simplifying the process and making sleep scoring a more attractive diagnostic tool. Below are the results of a neural network performing multiclass classification using the features listed in the table at left. (Regularization parameter λ=1.4) Time: Mean Median Maximum Kurtosis Variance Skewness # Zero Crossings Hjorth mobility and complexity Frequency: Factoral exponent Dominant freq Spectral edge freq Average freq Total bandpower Spectral moments Power ratios (α, β, γ, δ) Entropy: Spectral entropy Renyi entropy Relative spectral entropy Fisher information Accuracy = 80.82% Actual distribution:   W N1 N2 N3/N4 REM # per class 1828 573 3257 113 842 Predicted distribution: Sleep stage scoring criteria:   W N1 N2 N3/N4 REM 1649 35 100 20 172 80 236 63 174 27 3000 6 121 71 40 1 47 15 576 Stage Scoring Criteria (epoch length = 30 sec) W >50% alpha (8-13 Hz) or low voltage, mixed freq N1 <50% alpha activity; low voltage, mixed freq; slow eye movements and vertex sharp waves N2 Sleep spindles and K complexes; <20% high voltage (>75 µV), low freq (<2 Hz) N3/N4 >20% high voltage (>75 µV), low freq (<2 Hz) REM Low voltage, mixed freq; rapid eye movements; muscle atonia Neural Network Classification: a1(3) Feature Selection: Principal Component Analysis (PCA) allows for the dimension of the feature vector to be reduced without compromising accuracy: a1(2) x1 h(x) = g(z(3)) Design Overview: Preprocessing Feature Extraction Classification with Neural Network ah(3) xn ah(2) Feature Selection Preprocessing: g(x) is the activation of each unit in the network: Σg(x) All EEG time-series data were filtered using a tenth-order Butterworth bandpass filter in order to remove any frequency components outside the range of 0.1-64 Hz. The data was also subjected to a moving average filter and a notch filter at 60 Hz. Conclusion: The accuracy achieved in this project, 80.82 %, is a good start, but not ideal. This value can be improved with better feature extraction techniques, particularly those that take advantage of the characteristic waveforms found in N2. There is no doubt that a high performing automatic sleep scoring algorithm would be of tremendous value to the medical community. Training the neural network: Randomly initialize weights Θ(1), Θ(2) Implement forward propagation to compute h(x) Compute error using cost function Use back propagation to compute derivatives of cost fxn Increment weight values by cost function derivatives