Error-aware CNN improves automatic epileptic seizure detection

Automated seizure detection is a major challenge in the context of epilepsy diagnostics. There are numerous approaches to this task, but most of them share the same problem—the trade-off between recall and precision, i.e. decent recall is often accompanied by low precision. This ultimately leads to a high number of false positive seizure detections, which in its turn impede automated diagnostics. The purpose of this study is to develop a method to lower the number of false positive predictions in seizure detection task when applied to real EEG recordings. We propose the cascade approach which combines the idea of iterative refinement algorithms and powerful neural networks. The method is tested on unrefined dataset, that includes EEG recordings of epileptic patients from the hospital. Time-frequency analysis based on continuous wavelet transform is used for EEG preprocessing and feature extraction. To provide predictions the approach implements convolutional neural networks. The proposed approach consists of two steps: in the first step a model is trained to provide initial predictions and then in the second step another model is trained with the knowledge of the first model’s errors. We evaluate the performance of the approach with the confusion matrix metrics adjusted to the specifics of the epilepsy diagnostics task. We show that the number of false positive predictions decreases by an order of magnitude with the use of the proposed method. We theorize about possible application of this approach within a clinical decision support system.