Projects:2015s2-209 Automated Classification of Brain Activity During Sleep

Project Description

The project ‘Automated classification of brain activity during sleep’ has been assigned as a requirement to complete the master’s coursework in electronic engineering. The assigned project entails the group members to create a suitable sleep cycle classification to enable researchers to identify the sleep stages and use the data to further their research.

There are increasing number of cases where people are suffering from sleeplessness and chronic sleep ailments. Such ailments have to potential to jeopardise the lives of people as they are likely to lose alertness alongside weakening their immune system. These negative changes make them susceptible to various disorders. The most serious disorders that have been identified to affect the healthy mind and body are sleep apnoea and depression. These ailments are difficult to be corrected by a simple check-up.

Therefore, it is recommended to perform a sleep study to knowing the diseases affecting a healthy person before they further deteriorate the health. The project aims to capture the sleeping patterns spread out over the entire sleeping duration. The patterns are extracted from the brain sensors by an electroencephalogram (EEG). The EEG records the pattern to a European Data Format (EDF) file. The file becomes crucial in evaluating the factors that might be causing wakefulness. What is more, the anomalies in the sleeping pattern can be detected by comparing the EEG with a standard sleep cycle. The standard reference is available in an Extensive Mark-up Language (XML) format. The deviations of the actual sleep cycle from the ones displayed on the format are pivotal to capture and detect the sleep related signals.

The main objective of the project is to automatically classify the sleep stages of the data and detect any abnormality in the sleep cycle. Although the analysis of the time varying signal from the EEG may look out to be suitable to observe various sleep signals, the signal is time varying and would not be a true representation of the actual signals. Hence, the signal classification has to be done by implementing a suitable signal analysis approach to make the analysis of the data more meaningful to academicians.

Evidently, the most pivotal point of the project is the use of an appropriate signal analysis technique in order to have a high degree of accuracy in demonstrating the sleep stages. After taking into account many factors, signal processing is done by short term Fourier transformation (STFT) and wavelet transformation.It aids in extracting the critical components of the complex time varying signal making it easy to assess the frequency components. STFT enables us to break the lengthy time domain signal into smaller segments called windows, which can be subjected to fast Fourier transformation to simultaneously bring the frequency and amplitude information out of the signal. Even so, it has certain shortcomings considering that the window size is fixed. Therefore, wavelet analysis is used to eliminate these weaknesses. This evaluation gave certain features which were being trained by a neural network.

The final stage has been the training of the neural network in order to classify the sleep stages. These stages were automatically classified by the network. A high degree of accuracy was obtained while training the network. This can be attributed to the four layer neural network, which has given very precise results. Different records are tested and have been found to give a consistent result.

Project Results

- Using Wavelet Transform - Decompse the alpha, beta, theta and delta waves - Features Extraction

- ICA (Independent Component Analysis)

- Use Neural Network - Use Features as Input - Set the Target - Training

Project Conclusion

- For non-stationary signal, wavelet transform processing a good effect. - The ICA can remove most of the artifacts. - After using ICA, the accuracy of classification for sleep stages by using neural network will increase significantly.