Abstracts

Rationale:
Epilepsy is one of the most common neurological disorders. The diagnosis commonly requires manual visual electroencephalogram (EEG) analysis which is time-consuming. Deep learning has shown promising performance in detecting epileptiform discharges and may improve the quality of epilepsy monitoring. However, most of the datasets in the literature are small and collected in a single centre, limiting the generalizability of findings across different settings.
Method:
To study the performance of automated epileptic spike detection among different hospitals, we collected routine EEG recordings from patients with idiopathic generalized epilepsy (IGE) seen at the Alfred Hospital (n=94) and Royal Melbourne Hospital (RMH; n=115) hospitals in Melbourne, Australia. In addition, normal control recordings were obtained from these sites (n=98 and 120, respectively). We then trained a DeepIED model (Hao et al. 2018) and a deep learning Resnet for time series classification (Resnet-TSC) model (Fawaz et al. 2019) individually on these sets and cross-evaluated the models on sets from different hospitals. Band-pass filters of 0.5 - 50 Hz were applied to remove muscle artefacts. The EEG recordings were split into 2s segments containing either epileptiform discharges or normal background with 50% overlap. Windows from normal control EEG recordings were used as the normal segments in training steps to avoid misclassified epochs containing epileptiform discharges by neurologists. As it is common to have imbalanced datasets where there are more windows without epileptiform discharges, in respect of the Resnet-TSC, we studied 3 strategies to overcome this, oversampling, focal loss, and focal loss with oversampling. DeepIED solves this by learning the latent representation via triplet loss which was used as a benchmark for our approaches. We tested the models on another set of EEG recordings from Alfred (n=47) and RMH (n=56). Both test sets contained normal and abnormal EEG recordings, including background EEG from the abnormal recordings.
Results:
The results of classifying these windows are shown in Table 1 and Table 2. Regarding Resnet-TSC, focal loss and focal loss with oversampling achieved the best cross-evaluated area under the curve (AUC) scores on test sets from Alfred and RMH, 0.89 and 0.85, respectively. The DeepIED had slightly better results with the cross-evaluated AUC scores of 0.91 and 0.87 on the same test sets.
Conclusion:
Our proposed approaches to address the imbalanced dataset problem are able to achieve comparable results and generalize well across different IGE EEG datasets. Future work will aim to improve the models and collect more datasets from different hospitals with the ultimate goal of providing an inter-ictal epileptiform discharge detector that will be reliable across multiple settings and usable in the early stages of epilepsy diagnosis involving both routine and sleep-deprived EEG.
Funding:
:GRADUATE RESEARCH INDUSTRY PARTNERSHIPS SCHOLARSHIP