Code for a tool for the localization of epileptogenic zone based on full band features of stereo-EEG signal in patients with epilepsy, as described in the original paper: Localization of Epileptogenic Zone Using Full-band Features of SEEG.
This repository requires python3.6. Install the dependencies from PyPI:
pip3 install -r requirements.txtTo reproduce the major results reported in the paper, download the data from this link. put it under the data directory. Then simply run the notebook main_fig.ipynb.
If you want to see the original signal and time-frequency map of any channels by pressing on these three figures, you can also download and run another version with interaction:
python3 main.pyHigh frequency epileptogenicity index(HFEI) is a simplified version of epileptogenicity index(EI), which is a quantitative measure proposed by Bartolomei.The method combined spectral and temporal properties of high frequency activities during seizure onset.
First, a 35-senconds signal was chosen and a band-pass FIR was used to obtain signals in the gamma and high-gamma of both ictal and baseline. Then, the band-passed SEEG signal was transformed into the high-frequency energy by amplitude squaring and window smoothing for each channel. The ictal high-frequency energy was normalized by dividing the average value of high-frequency energy of the baseline. The normalized high-frequency energy was shown below, where x-axis represents time, y-axis represents channels and color represents the energy:
We calculate time coefficient (TC) and energy coefficient (EC) according to the normalized high frequency energy (NHFE). TC of each electrode was defined as the reciprocal of order of the onset time Nk. The higher TC represented the earlier onset of the channel. Taking the first Nk as the seizure onset time N0, EC was defined as the normalized high frequency energy during a 250ms period after N0. The higher the EC, the higher the high-frequency energy of the channel at seizure onset. We supposed the channels inside the EZ has earlier onset (TC) and higher energy (EC) of high-frequency activity at the onset. HFEI of each channel was calculated as the product of TC and EC.The HFEI was shown below, where x-axis represents channel, y-axis represents the value of HFEI, and the red dotted line represents the threshold, which was set to mean(HFEI)+std(HFEI)
The 10 channels with the highest HFEI were selected to form a 'committee', which voted for the epileptic cluster in next step.
In addition to high-frequency activity, there are other potential biomarkers of epileptogenic zone, such as low-frequency suppression and spikes. We used an unsupervised learning method to distinguish time-frequency patterns inside the epileptogenic zone from those outside the epileptogenic zone.
The time-frequency pattern was extracted by calculating the time-frequency map for each channel first. Then the time-frequency map, containing 427500 pixels, was then expanded into a vector.
The exorbitant dimension would bring a huge amount of computation to the computer, and also redundant information or noise information, which may result in overfitting. Therefore, it is necessary to reduce the dimension. Here we used principal component analysis(PCA) as the method for dimension reduction.
After the the dimension reduction, channels would be transformed into the PC space, in which two points closer together tend to have similar time-frequency pattern. We used a distance-based clustering method, K-means, to categorize the channels.
After clustering, channels were divided into different clusters. The 'committee' made up by the 10 channels with the highest HFEI voted for the epileptogenic cluster. In the figure below, the orange dots represents the channels in the 'committee', the yellow/orange dots with text annotation represents the channels within the epileptogenic cluster.
If you have a question or feedback, or find any problems, you can contact us by email or open an issue on GitHub.