Cyclops: Automated detection of interictal epileptiform discharges with few EEG channels

Abstract

Interictal epileptiform discharges (IEDs) are crucial for epilepsy diagnosis and management. New EEG devices with fewer electrodes are more accessible but their ability to detect IEDs is uncertain. The aim of this study is to develop and validate a machine learning model capable of detecting IEDs in reduced-channel EEG data, enabling broader epilepsy diagnosis.

Using EEG samples from 3,378 patients and an external validation set of 51 patients, we trained Cyclops, a deep neural network designed to function across various channel configurations.

Performance was evaluated using AUROC and other clinically relevant metrics, including IED source location sensitivity. Cyclops demonstrated strong performance even with minimal channels. AUROC for one channel: 0.876 [95% CI: 0.854-0.897]; best configuration based on a clinically available product: 0.950 [95% CI: 0.936-0.962]; for the detection of focal IEDs with two local channels, AUROC values ranged from 0.701 [95% CI: 0.656-0.745] to 0.930 [95% CI: 0.902-0.955] with a median AUROC of 0.809. On the external validation set, performance ranged from 0.692 [95% CI: 0.593-0.782] to 0.949 [95% CI: 0.922-0.972] with a median AUROC of 0.846. Thus, Cyclops supports effective IED detection with reduced EEG setups, enhancing accessibility and expanding epilepsy diagnosis to broader patient populations.

Background

Interictal epileptiform discharges (IEDs) are transient electrical abnormalities detected in electroencephalograms (EEGs) and are essential for epilepsy diagnosis. Traditional EEG setups require a full 10-20 electrode montage, which is resource-intensive and limits accessibility, particularly in remote or underserved regions. Reduced-channel EEG devices are more portable and easier to use but have not been thoroughly validated for detecting IEDs. Previous research has mainly focused on seizure detection or standard full-montage IED analysis, leaving a gap in understanding the feasibility of reduced-channel EEGs for this purpose. This study aims to bridge this gap by developing a machine learning model capable of accurately identifying IEDs using fewer electrodes, potentially expanding diagnostic capabilities to broader patient populations.

Methods

The study used EEG data from 3,631 patients, including an internal dataset (Massachusetts General Hospital) and an external validation set (Human Epilepsy Project). A deep convolutional neural network (Cyclops) was trained to predict IED probability across different EEG montages, with a primary split (80/10/10 for training, validation, and testing) ensuring robust evaluation. Data augmentation techniques, including jittering, amplitude scaling, and mirroring, improved generalization. Model performance was assessed using the area under the receiver operating characteristic curve (AUROC) across multiple channel configurations. Statistical significance was determined through bootstrapping, and performance was validated on an independent dataset to confirm generalizability.

Data Description

The internal dataset consisted of 112,154 EEG samples from 3,580 patients recorded at Massachusetts General Hospital. These samples were categorized into three subsets: (1) Internal A (18,804 samples) with expert-annotated IEDs, (2) Internal B (63,350 samples) containing pseudo-labeled IEDs based on morphological similarity, and (3) Internal C(30,000 negative samples) drawn from EEGs without IEDs. Data were split into 80/10/10 for training, validation, and testing. IED localization was determined by consensus among multiple neurologists, with training data including augmentation techniques to enhance model robustness. The test set contained only high-confidence IED samples to ensure reliable evaluation.

Usage Notes

The data for training Cyclops is provided here in this repository. Code for training Cyclops is available in the associated GitHub repository (https://github.com/bdsp-core/cyclops). 

Ethics

This project was conducted under protocols approved by the BIDMC IRB. Protocol numbers: 2024P000804, 2022P000481, 2022P000417

This dataset does not include any PHI. 

Conflicts of Interest

MBW was supported by grants from the NIH (RF1AG064312, RF1NS120947, R01AG073410, R01HL161253, R01NS126282, R01AG073598, R01NS131347, R01NS130119), and NSF (2014431). He is a co-founder, scientific advisor, and consultant to Beacon Biosignals and has a personal equity interest in the company. DMG was supported by NIH K23NS124656. JJH is supported by the VA ORD (I01HX003107-01A2) and the Human Epilepsy Project. MG is supported by the SC BIDS4Health training grant from the National Institute of Health (T15LM013977). DD is on the Scientific Advisory Board for Beacon Biosignals. EG was supported by awards UL1TR002378 and KL2TR002381 (NCATS/NIH) as well as R01NS110347 (NINDS/NIH). JY is supported by the NIH 1UH3NS109557-01A1 and receives royalty from Elsevier.  MG is supported by funding from UNEEG.

Dr. Westover is a co-founder, scientific advisor, consultant to, and has personal equity interest in Beacon Biosignals.