Towards Long-term Non-invasive Monitoring for Epilepsy via Wearable EEG Devices

Abstract

We present the implementation of seizure detection algorithms based on a minimal number of EEG channels on a parallel ultra-low-power embedded platform. The analyses are based on the CHB-MIT dataset, and include explorations of different classification approaches (Support Vector Machines, Random Forest, Extra Trees, AdaBoost) and different pre/postprocessing techniques to maximize sensitivity while guaranteeing no false alarms. We analyze global and subject-specific approaches, considering all 23-electrodes or only 4 temporal channels. For 8 s window size and subject-specific approach, we report zero false positives and 100% sensitivity. These algorithms are parallelized and optimized for a parallel ultra-low power (PULP) platform, enabling 300h of continuous monitoring on a 300 mAh battery, in a wearable form factor and power budget. These results pave the way for the implementation of affordable, wearable, long-term epilepsy monitoring solutions with low falsepositive rates and high sensitivity, meeting both patient and caregiver requirements.

Energy-Efficient Tree-Based EEG Artifact Detection

“…Implementing an artifact detecting Extra Trees model on a multi-core edge platform (Mr. Wolf) requires a power envelope of only ≈22 mW with sub-200 µs processing time, thus with much lower energy re- quirements than the analog front-end and BLE of the complete platform that enables multi-day functionality [23]. In combination with previous results for low energy implementation of seizure detection on Mr. Wolf [21], these pave the way to the design of robust and unobtrusive EEG wearable devices.…”

“…Furthermore, we deployed and optimized the algorithms on a PULP platform, achieving minimal energy requirements (≈ 4 µJ per inference), outperforming competing commercial devices. These results, combined with robust epilepsy detection models [21], show that a PULP system is indeed one of the best candidates for future wearable epilepsy monitoring systems based on minimal EEG setups. Future work will focus on integrating the proposed artifact detection model with a robust epilepsy detection framework and on expanding the proposed techniques to include sensor fusion from other data sources, to be realized in a wearable setting of a body area network and on testing the whole system in ambulatory and domestic environments.…”

“…DWT is a signal decomposition technique for timefrequency analyses [26] that has been reported to work well for seizure detection [21]. DWT offers a good trade-off between signal-to-noise ratio and performance [27], and it allows to obtain both frequency and temporal features of a signal by applying a cascade series Mother Wavelets filters.…”

“…Following the approach of [21], we implement the artifact detection framework on the BioWolf wearable ExG device [23] (Fig. 2).…”

“…Within this context, we present a framework to automatically detect EEG artifacts, designed to aid and support more robust epilepsy detection systems [21] for wearable unobtrusive epilepsy detection with a reduced montage. Specifically, the main contributions are the following:…”