Minimizing artifact-induced false-alarms for seizure detection in wearable EEG devices with gradient-boosted tree classifiers

www.nature.com/scientificreports OPEN Minimizing artifact‑induced false‑alarms for seizure detection in wearable EEG devices with gradient‑boosted tree classifiers Thorir Mar Ingolfsson 1*, Simone Benatti 2,3, Xiaying Wang 1, Adriano Bernini 4, Pauline Ducouret 4, Philippe Ryvlin 4, Sandor Beniczky 5,6, Luca Benini 1,2 & Andrea Cossettini 1 Electroencephalography (EEG) is widely used to monitor epileptic seizures, and standard clinical practice consists of monitoring patients in dedicated epilepsy monitoring units via video surveillance and cumbersome EEG caps. Such a setting is not compatible with long-term tracking under typical living conditions, thereby motivating the development of unobtrusive wearable solutions. However, wearable EEG devices present the challenges of fewer channels, restricted computational capabilities, and lower signal-to-noise ratio. Moreover, artifacts presenting morphological similarities to seizures act as major noise sources and can be misinterpreted as seizures. This paper presents a combined seizure and artifacts detection framework targeting wearable EEG devices based on Gradient Boosted Trees. The seizure detector achieves nearly zero false alarms with average sensitivity values of 65.27% for 182 seizures from the CHB-MIT dataset and 57.26% for 25 seizures from the private dataset with no preliminary artifact detection or removal. The artifact detector achieves a state-of-the-art accuracy of 93.95% (on the TUH-EEG Artifact Corpus dataset). Integrating artifact and seizure detection significantly reduces false alarms—up to 96% compared to standalone seizure detection. Optimized for a Parallel Ultra-Low Power platform, these algorithms enable extended monitoring with a battery lifespan reaching 300 h. These findings highlight the benefits of integrating artifact detection in wearable epilepsy monitoring devices to limit the number of false positives. Epilepsy is a common neurological disorder that affects more than 50 million people w orldwide1 and is characterized by the recurrence of seizures which temporarily compromise the function of the affected people’s brain. About one-third of persons with epilepsy (PWE) continue to suffer seizures despite receiving appropriate antiseizure medications. In most instances, these persons will lose awareness during their seizures, putting them at risk of accidents, traumatism, and even death. Furthermore, PWE will often fail to remember the occurrence of their seizures and will thus be unable to inform their physician to adjust therapy appropriately. These issues have led to an increasing interest in developing seizure detection solutions with two goals: (1) sending an alarm to family members or caregivers to protect PWE from the immediate risks entailed by seizures, (2) providing a reliable seizure count that they can share with their physicians to optimize treatment. Several methods are being developed for this purpose; most currently rely on biosignals captured at the wrist or arm, including surface electromyography, 3D-accelerometry, electrodermal activity, and photoplethysmography. While several methods proved effective in detecting one subtype of seizures, i.e., generalized tonic-clonic seizures (GTCS), they fall short of detecting most other seizure types. For this reason, other methods attempt to detect seizures using the classic neurobiological hallmark of epileptic seizures, Electroencephalography (EEG). Indeed, by definition, a seizure reflects an abnormal EEG signal called an epileptic discharge. Diagnostic methods 1 ETH Zürich, D-ITET, 8092 Zürich, Switzerland. 2University of Bologna, 40126 Bologna, Italy. 3University of Modena and Reggio Emilia, 41121 Reggio Emilia, Italy. 4University Hospital of Lausanne (CHUV), 1011 Lausanne, Switzerland. 5Aarhus University Hospital, 8200 Aarhus, Denmark. 6Danish Epilepsy Centre (Filadelfia), 4293 Dianalund, Denmark. *email: Scientific Reports | (2024) 14:2980 | https://doi.org/10.1038/s41598-024-52551-0 1 Vol.:(0123456789) www.nature.com/scientificreports/ to perform short-term EEG recordings (i.e., from 20 min to several days) are very well established but not adapted to the purpose of chronic recordings over months or years to achieve the above-described objectives. Indeed, conventional EEG systems are bulky and uncomfortable, causing patients to perceive stigmatization. Furthermore, the long wires used to connect multiple electrodes are a significant cause of motion artifacts on the EEG traces2. Consequently, wearable solutions are paramount for long-term continuous EEG monitoring3. In this context, the need for wearables-based long-term seizure detection requires empowering such solutions with seizure-detecting capabilities via Machine Learning (ML), to enable prompt interventions from caregivers during or immediately after the seizures, reducing their impact and providing more reliable information to the physicians (to optimize anti-seizure therapies)4. However, the development of EEG-based seizure detectors for wearable Internet of Things (IoT) devices is faced by multiple challenges, which we address in this paper. First, most of the existing Artificial Intelligence (AI) models rely on a large number of electrodes5 (ideally all electrodes of standard EEG-caps): unobtrusive wearable solutions are limited to a lower number of channels, and they face the challenge of maintaining the same levels of performance as for full-channel systems. In addition, the impact of false alarms is much greater in long-term monitoring settings (since it relates to the willingness of patients to use the devices), ultimately resulting in the need to maximize specificity (also at the price of a lower sensitivity) as the main performance m etric6. Second, artifacts play a critical role in wearable IoT devices. While data acquired in epilepsy monitoring units (EMUs) usually are inspected by experts and feature a low amount of artifacts (which can be labeled)7, wearable EEG systems produce signals with low signal-to-noise ratio8 that are more affected by artifacts than full EEG-caps or implanted solutions. If not accounted for, artifacts can significantly increase the number of false alarms9, possibly making a wearable seizure detector not usable in practical settings (low specificity). Therefore, automated seizure detection frameworks must be combined with artifact detection (and, possibly, filtering). Last but not least, wearable IoT devices must also fulfill the following requirements: (a) small and comfortable form factor; (b) long-battery life; (c) low latency. To ensure this, smart edge computing based on low-power microcontrollers (MCUs) has recently been introduced, proving its effectiveness in providing long-term operation and executing AI m odels10. However, the challenge is that the AI algorithms need to fit the computational capabilities of wearable devices. Consequently, the choice of models has to be narrowed down, so that they can be impleme (...truncated)