The time series captured by a single scalp electrode (plus the reference electrode) of refractory epileptic patients is used to forecast seizures susceptibility. The time series is preprocessed, segmented, and each segment transformed into an image, using three different known methods: Recurrence Plot, Gramian Angular Field, Markov Transition Field. The likelihood of the occurrence of a seizure in a future predefined time window is computed by averaging the output of the softmax layer of a CNN, differently from the usual consideration of the output of the classification layer. By thresholding this likelihood, seizure forecasting has better performance. Interestingly, for almost every patient, the best threshold was different from 50%. The results show that this technique can predict with good results for some seizures and patients. However, more tests, namely more patients and more seizures, are needed to better understand the real potential of this technique.
Refractory epileptic patients can suffer a seizure at any moment. Seizure prediction would substantially improve their lives. In this work, based on scalp EEG and its transformation into images, the likelihood of an epileptic seizure occurring at any moment is computed using an average of the softmax layer output (the likelihood) of a CNN, instead of the output of the classification layer. Results show that by analyzing the likelihood and thresholding it, prediction has higher sensitivity or a lower FPR/h. The best threshold for the likelihood was higher than 50% for 5 patients, and was lower for the remaining 36. However, more testing is needed, especially in new seizures, to better assess the real performance of this method. This work is a proof of concept with a positive outlook.