Using a convolutional neural net to predict epilepsy diagnoses from patient EEG recordings.
The CNN is trained on EEG data from the Temple University EEG Corpus, which is a collection of EEGs freely available to the public. We use a subset of EEGs from this corpus from patients with a diagnosis of either epilepsy or no epilepsy. Data from patients diagnosed with epilepsy and patients diagnosed with no epilepsy are contained in two separate directories and organized by patient ID. A more in depth description of the dataset can be found here.
View the code for data wrangling here.
There are a number of different ways information can be extracted from EEG signal data to train a CNN. The most common method for epilepsy detection is to train on raw signals (Craik et al., 2019), so I chose to adopt this method.
The EEG in this dataset was recorded with variable parameters.
- Sampling Rate: 250, 256, 512
- Reference: Average, Linked Mastoid
- Number of Channels: 24, 26
In order to keep the input parameters consistent across all instances fed into the CNN, I chose to only use EEGs recorded with a 250hz sampling rate. This gave me a pretty even split between EEGs with epilepsy and EEGs with no epilepsy. I also chose to only use EEGs with an average reference. For the EEGs that met these criteria, I chose to only use signal data from the 24 channels shared by all recordings in the dataset. These channels were:
Fp1-REF, Fp2-REF, F3-REF, F4-REF, C3-REF, C4-REF, P3-REF, P4-REF, O1-REF, O2-REF, F7-REF, F8-REF, T3-REF, T4-REF, T5-REF, T6-REF, M1-REF, M2-REF, FZ-REF, CZ-REF, PZ-REF, ECG1-REF, T1-REF, T2-REF
Because these EEGs have different durations, I selected a duration of 10 seconds (2500 samples if recorded at 250hz) to feed into the CNN. Epilepsy is typically diagnosed from EEG by the observation of approximately 3hz spike waves that usually last between 5 and 10 seconds (Smith, 2005). I reasoned that 10 seconds of data would be able to catch these spikes if they are present. For each EEG recording that met the criteria above, I split the recording into 10 second segments and counted each of these segments as a separate instance.
Applying a 0.9, 0.1 split to the data, I ended up with the following counts
| Epilepsy | No Epilepsy | Total | |
|---|---|---|---|
| Train | 25812 | 19154 | 44966 |
| Test | 2876 | 2121 | 4997 |
View the code for training the CNN here.
I chose a 1-Dimensional CNN to train on this data, following the example of Ihsan et al., 2018. A 1D CNN works well for training on time-series data, where signals recorded from different channels are not necessarily related. I used the example in this Ackermann, 2018 article as a template for constructing the neural net. I added four convolutional layers and a final dense layer that predicts liklihood values for the two possible outputs (1=Epilepsy, 0=No Epilepsy).
| Layer Type | Filters | Kernel Size | Output Shape | Param # |
|---|---|---|---|---|
| Conv1D | 100 | 10 | (None, 2491, 100) | 24100 |
| Conv1D | 100 | 10 | (None, 2482, 100) | 100100 |
| MaxPooling1D | (None, 827, 100) | 0 | ||
| Conv1D | 160 | 10 | (None, 818, 160) | 160160 |
| Conv1D | 160 | 10 | (None, 809, 160) | 256160 |
| GlobalAveragePooling1D | (None, 160) | 0 | ||
| Dropout | (None, 160) | 0 | ||
| Dense | (None, 2) | 322 |
I trained the CNN using a validation split of 0.1 and 50 training epochs. Interestingly, for all 50 epochs, the model showed a training accuracy of 0.5736 and a validation accuracy of 0.5773, showing that multiple training epochs did not improve the accuracy. Evaluating the model on the testing data gave an accuracy score of 0.5763.
This first pass was largely a test to see if I could train a CNN with this data and get some kind of predictions out of it. While I accomplished the goal of training a model, the trained model did not have very much predictive power for any of the datasets that were fed in (train, val, or test). There are a number of different parameters I can tweak to achieve better predicive power. Because it tends to have large baseline shifts that obscure the signal, EEG signals are often highpass filtered at around 0.1hz. I am not sure if these data were previously highpass filtered, but if it was not, this could have greatly reduced the neural net's ability to pick out subtle features in the signal. Additionally, there could have been certain channels that were not providing helpful information for epilepsy diagnosis. It would be benefitial to research which scalp locations are most likely to pick up epileptic spikes and only train on channels close to these locations. I could have easily attained a much larger dataset if I figured out a way to downsample recordings with a higher sampling rate than 250hz. In order to generalize the use of thie CNN to as many EEG recordings as possible, I could downsample all recordings to an even lower sampling rate (100hz for instance) before feeding in the signals to the CNN. As far as the neural net itself, there are countless parameters I could have changed or different structures I could have used. The 1D CNN seemed to be the best type of CNN for time series data, but besides that I did not have much guidance on how to structure it. A good future direction would be to look at how Ihsan et al., 2018 structured their CNN for predicting epilepsy diagnoses. They claim 99.1±0.9% accuracy for their model and give a fairly detailed description of the structure. Even though this first pass was more of an experiment of what would be possible as far as training a neural network on EEG data, I now have a good framework going forward for testing out different models and parameters on this dataset.
Craik, A., He, Y., & Contreras-Vidal, J. L. (2019). Deep learning for electroencephalogram (EEG) classification tasks: A review. Journal of Neural Engineering, 16(3). https://doi.org/10.1088/1741-2552/ab0ab5
Smith, S. J. M. (2005). EEG in the diagnosis, classification, and management of patients with epilepsy. Neurology in Practice, 76(2). https://doi.org/10.1136/jnnp.2005.069245
Ullah, I., Hussain, M., Qazi, E. ul H., & Aboalsamh, H. (2018). An automated system for epilepsy detection using EEG brain signals based on deep learning approach. Expert Systems with Applications, 107, 61–71. https://doi.org/10.1016/j.eswa.2018.04.021