Epileptic seizure forecasting with long short-term memory (LSTM) neural networks
Authors:
Daniel E. Payne,
Jordan D. Chambers,
Anthony Burkitt,
Mark J. Cook,
Levin Kuhlman,
Dean R. Freestone,
David B. Grayden
Abstract:
Objective: Forecasting epileptic seizures can reduce uncertainty for patients and allow preventative actions. While many models can predict the occurrence of seizures from features of the EEG, few models incorporate changes in features over time. Long Short-Term Memory (LSTM) neural networks are a machine learning architecture that can display temporal dynamics due to the recurrent connections. In…
▽ More
Objective: Forecasting epileptic seizures can reduce uncertainty for patients and allow preventative actions. While many models can predict the occurrence of seizures from features of the EEG, few models incorporate changes in features over time. Long Short-Term Memory (LSTM) neural networks are a machine learning architecture that can display temporal dynamics due to the recurrent connections. In this paper, we used LSTMs to monitor changes in EEG features over time to improve the accuracy of seizure forecasts and to alter the time window of the forecast. Methods: Long-term intracranial EEG recordings from eight patients from the NeuroVista dataset were used. A Fourier transform of 1-minute segments of EEG was fed into a Convolutional Neural Network (CNN). The outputs from the CNN were input to three different LSTM models at different time intervals: 1 minute, 1 hour and 1 day. The LSTM model outputs were used to predict seizure onset within a time window. The prediction and start of the time window were separated by the same length of time as the window. Window sizes tested included 2, 4, 10, 20 and 40 minutes. Results and Conclusion: Our model forecast seizure onsets well above a random predictor. Compared to other models using the same dataset, our model performed better for some patients and worse for others. Monitoring the change in EEG features over time allowed our model to produce good results over a range of different window sizes, which is an improvement on previous models and raises the possibility of altering the forecast to meet individual patient needs. Furthermore, a window size of 40 minutes provides a potential intervention time of 40 minutes, which is the first time an intervention time of more than 5 minutes have been forecast using long-term EEG recordings.
△ Less
Submitted 18 September, 2023;
originally announced September 2023.
