Rationale: Epilepsy is an intriguing dynamic brain disorder whereby epileptic seizures are occurring roughly 0.05% of the time. Yet, they are associated with a high potential for disability or even death. Many people with epilepsy continue to have seizures despite medication therapy, surgical treatments, and neuromodulation therapy. The unpredictability of seizures is one of the most disabling aspects of epilepsy, leading to significant interest in forecasting seizures to help patients manage activities or facilitate targeted therapies. EEG has been the modality most studied for the development of seizure forecasting algorithms. However, major limitations to forecasting have been the scarcity of long-term EEG data, except from invasive risk-prone intracerebral recordings. New subcutaneous continuous EEG recording systems have shown promise to advance diagnosis and treatment in epilepsy.
Methods: We used the longest subcutaneous EEG (sqEEG) recording to date, over 230 days in length, to develop an algorithm capable of forecasting seizures from this novel data source as the capability of predicting seizures with subcutaneous EEG is still not known.
We explored several network modifications and preprocessing techniques to design a seizure prediction algorithm using LSTM deep learning classifier.
First, electrographic seizures were identified by visual review of a trained epileptologist. Based on confirmed seizures, data was labeled as preictal (one-hour data segments with a set-back of 5 minutes before seizure onset) or interictal segments (seizure-free periods at least one day apart from any seizure). Next, the data was segmented into one-minute epochs and preprocessed. To compensate for the heavily unbalanced data ratio in training, noise-added copies of preictal data segments were generated. Finally, the entire data set was normalized (z-score).
The configurations evaluated are:
Configurations 1-2: Training input data included two filtered channels and 2 FFT channels trained in the architecture of 3 LSTM layers with 200/100 units each.
Configurations 3-5: Training input data included two filtered and down-sampled channels, 2 FFT channels, and one time of day channel trained in the following architectures:
- Configuration 3: 10 LSTM with 25 units in each layer architecture.
- Configuration 4: 2 BiLSTM with 20 units in each layer architecture.
- Configuration 5: SimpleRNN architecture.
Results: The area under the receiver operating characteristic (ROC) curve (AUC) was calculated by taking the average probability for 5-min segments per label, then the maximum probability of each hour (12 values of 5-min segments) to evaluate the different configurations. Results are shown in table 1.
Conclusions: These preliminary results suggest that it is possible to forecast seizures using two-channel chronic subcutaneous EEG recordings. Future work will be focused on improving the sensitivity and specificity of seizure forecasts and extending results to additional patients currently recording data in the home environment.
Funding: Please list any funding that was received in support of this abstract.: This work was funded by the ‘My Seizure Gauge’ grant provided by the Epilepsy Innovation Institute, a research program of the Epilepsy Foundation of America.