Objective Automated detection of interictal epileptiform discharges (IEDs) on electroencephalographic (EEG) data aims to reduce the time and resources spent on visual analysis by experts (the gold standard) with algorithms that match or outperform experts. In this study, we aimed
...
Objective Automated detection of interictal epileptiform discharges (IEDs) on electroencephalographic (EEG) data aims to reduce the time and resources spent on visual analysis by experts (the gold standard) with algorithms that match or outperform experts. In this study, we aimed to further improve IED detection performance of a deep neural network based algorithm with a simpler second-level postprocessing deep learning network, a new approach in this field. Materials and Methods Seventeen interictal ambulatory EEGs were used, 15 with focal and 2 with generalized epilepsy in patients of aged 4 to 80 years (median: 19 years; 25th-75th percentile: 14-32 years). Two-second nonoverlapping epochs with a 0.99 or higher IED probability were selected by a previously developed VGG-C convolutional neural network (CNN) as input for the second-level postprocessing CNN we developed. Our CNN was tested on the resulting 580 EEG epochs after 80/20 training/validation with 3,049 epochs. Results Model accuracy was 86% for the validation set and 60% for the test set. The first-level CNN selected 37% true IEDs, and with the addition of our second-level postprocessing CNN, this increased to 38%. Doubling input data of the second-level CNN, and making its architecture more complex, as well as less complex, did not improve performance. Conclusion We were unable to reproduce the previously reported performance of the first-level CNN, and adding the postprocessing CNN did not improve IED detection.
