Epilepsy is the most common neurological disorder. The diagnosis commonly requires manual visual electroencephalogram (EEG) analysis which is time-consuming. Deep learning has shown promising performance in detecting interictal epileptiform discharges (IED) and may improve the quality of epilepsy monitoring. However, most of the datasets in the literature are small (n≤100) and collected from single clinical centre, limiting the generalization across different devices and settings. To better automate IED detection, we cross-evaluated a Resnet architecture on 2 sets of routine EEG recordings from patients with idiopathic generalized epilepsy collected at the Alfred Health Hospital and Royal Melbourne Hospital (RMH). We split these EEG recordings into 2s windows with or without IED and evaluated different model variants in terms of how well they classified these windows. The results from our experiment showed that the architecture generalized well across different datasets with an AUC score of 0.894 (95% CI, 0.881–0.907) when trained on Alfred’s dataset and tested on RMH’s dataset, and 0.857 (95% CI, 0.847–0.867) vice versa. In addition, we compared our best model variant with Persyst and observed that the model was comparable.

No abstract

Objective: Automated interictal epileptiform discharge (IED) detection has been widely studied, with machine learning methods at the forefront in recent years. As computational resources become more accessible, researchers have applied deep learning (DL) to IED detection with promising results. This systematic review aims to provide an overview of the current DL approaches to automated IED detection from scalp EEG and establish recommendations for the clinical research community. Methods: We conduct a systematic review according to the PRISMA guidelines. We searched for studies published between 2012 and 2021 implementing DL for automating IED detection from scalp EEG in major medical and engineering databases. We highlight trends and formulate recommendations for the research community by analyzing various aspects: data properties, preprocessing methods, DL architectures, evaluation metrics and results, and reproducibility. Results: The search yielded 67 studies, and 17 met our inclusion criteria. There were 2 main DL networks, convolutional neural networks in 14 studies and the long short-term memory networks in 3 studies. A hybrid approach combining a hidden Markov model with an autoencoder was employed in 1 study. All DL models involved supervised learning. The median number of layers was 11 (IQR: 5 - 19). The median number of IEDs was 10,122 (IQR: 2,757 - 16,824). Only 6 studies acquired data from multiple clinical centers. AUC was the most reported metric (median: 0.94; IQR: 0.92 - 0.96). Significance: The application of DL to IED detection is still limited and lacks standardization in data collection, multi-center testing, and reporting of clinically relevant metrics (i.e., F1, AUCPR, and false-positive/minute). However, the performance is promising, suggesting that DL might be a helpful approach. Further testing on multiple datasets from different clinical centers is required to confirm the generalizability of these methods.

Background and Objectives:To examine the preferences and user experiences of people with epilepsy and caregivers regarding automated wearable seizure detection devices.Methods:We performed a mixed-methods systematic review. We searched electronic databases for original peer-reviewed publications between January 1, 2000, and May 26, 2021. Key search terms included “epilepsy”, “seizure”, “wearable”, and “non-invasive”. We performed a descriptive and a qualitative thematic analysis of the studies included according to the technology acceptance model. Full texts of the discussion sections were further analyzed to identify word frequency and word mapping.Results:Twenty-two observational studies were identified. Collectively, they comprised responses from 3299 participants including patients with epilepsy, caregivers and healthcare workers. Sixteen studies examined user preferences, five examined user experiences, and one examined both experiences and preferences. Important preferences for wearables included improving care, cost, accuracy, and design. Patients desired real-time detection with a latency of ≤15 minutes from seizure occurrence, along with high sensitivity (≥90%) and low false-alarm rates. Device related costs were a major factor for device acceptance, where device costs of <$300 USD and a monthly subscription fee of <$20 USD were preferred. Despite being a major driver of wearable-based technologies, sudden unexpected death in epilepsy (SUDEP) was rarely discussed. Among studies evaluating user experiences, there was a greater acceptance towards wristwatches. Thematic coding analysis showed that attitudes towards device use, and perceived usefulness were reported consistently. Word mapping identified ‘specificity’, ‘cost’, and ‘battery’ as key single terms, and ‘battery life’, ‘insurance coverage’, ‘prediction/detection quality’, and the effect of devices on ‘daily life’ as key bigrams.Discussion:User acceptance of wearable technology for seizure detection was strongly influenced by accuracy, design, comfort, and cost. Our findings emphasise the need for standardised and validated tools to comprehensively examine preferences and user experiences of wearable devices in this population, using the themes identified in this study. Greater efforts to incorporate perspectives and user experiences in developing wearables for seizure detection, particularly in community-based settings are needed.PROSPERO Registration:CRD42020193565.