Name: SpikeNet 2.0
Published: April 30, 2025
License: https://github.com/bdsp-core/bdsp-license-and-dua

Database Restricted Access

Jun Li , Daniel Goldenholz , Shenda Hong , Chenxi Sun , Jin Jing , M Brandon Westover

Published: April 30, 2025. Version: 1.0

When using this resource, please cite: (show more options)
Li, J., Goldenholz, D., Hong, S., Sun, C., Jing, J., & Westover, M. B. (2025). SpikeNet 2.0 (version 1.0). Brain Data Science Platform. https://doi.org/10.60508/mbxb-hn49.

MLA	Li, Jun, et al. "SpikeNet 2.0" (version 1.0). Brain Data Science Platform (2025), https://doi.org/10.60508/mbxb-hn49.
APA	Li, J., Goldenholz, D., Hong, S., Sun, C., Jing, J., & Westover, M. B. (2025). SpikeNet 2.0 (version 1.0). Brain Data Science Platform. https://doi.org/10.60508/mbxb-hn49.
Chicago	Li, Jun, Goldenholz, Daniel, Hong, Shenda, Sun, Chenxi, Jing, Jin, and M Brandon Westover. "SpikeNet 2.0" (version 1.0). Brain Data Science Platform (2025). https://doi.org/10.60508/mbxb-hn49.
Harvard	Li, J., Goldenholz, D., Hong, S., Sun, C., Jing, J., and Westover, M. B. (2025) 'SpikeNet 2.0' (version 1.0), Brain Data Science Platform. Available at: https://doi.org/10.60508/mbxb-hn49.
Vancouver	Li J, Goldenholz D, Hong S, Sun C, Jing J, Westover M B. SpikeNet 2.0 (version 1.0). Brain Data Science Platform. 2025. Available from: https://doi.org/10.60508/mbxb-hn49.

Abstract

SpikeNet2 is an advanced deep learning model for automated detection of epileptiform discharges (spikes) in electroencephalogram (EEG) recordings. The model achieves expert-level performance in both event-level spike detection and EEG-level classification, with significantly reduced false positive rates compared to previous models. Trained on a large multi-center dataset of 17,812 EEGs from 13,523 patients and validated across multiple external datasets, SpikeNet2 demonstrates high generalizability and clinical utility for epilepsy diagnosis and monitoring. Code is provided here: https://github.com/bdsp-core/SpikeNet2](https://github.com/bdsp-core/SpikeNet2

Background

Electroencephalography (EEG) is the primary diagnostic tool for evaluating epilepsy, with epileptiform discharges serving as crucial biomarkers. However, manual interpretation of EEGs requires specialized expertise, which is limited globally. Previous AI models have focused on either event-level detection or EEG-level classification, but not both simultaneously. Additionally, high false positive rates have limited clinical application of automated detection systems. SpikeNet2 was developed to address these limitations through a novel hard negative mining approach and comprehensive dual-level classification.

Methods

This study of human subjects was approved by the Mass General Brigham Institutional Review Board (IRB approval # 2012P001929), including review of EEG and other clinical data. The Partners Healthcare Human Research Committee provided a waiver of written consent for this study. All data is deidentified.

We analyzed 17,812 EEGs from 13,523 patients across three datasets (MGB, HEP, and SCORE-AI). Event-level labeling was conducted by 24 experts from 18 institutions on 32,433 events from 2,601 patients. We developed a ResNet-based model incorporating an attention mechanism and implemented hard negative mining to reduce false positives. Performance was assessed using AUC-ROC, AUC-PR, calibration error, and a "modified" ROC metric.

Data Description

The dataset includes 17,524 EEGs from Massachusetts General Brigham hospitals (MGH and BWH), 188 EEGs from the Human Epilepsy Project, and 100 EEGs from the SCORE-AI dataset. Patient ages ranged from newborns to over 100 years (median: 53 years), with 47% female participants. EEGs were recorded in outpatient (58%), epilepsy monitoring unit (6%), and long-term monitoring (34%) settings. The data includes comprehensive expert annotations, with event-level samples receiving between 1-23 votes each (median: 4 votes).

Usage Notes

Data and code to generate all results and figures from the publication are provided here. The repository includes the trained models, processing pipelines, and analysis scripts required to reproduce the findings. Additionally, the complete dataset is available on AWS (see below - requires signing the DUA) to facilitate benchmarking of future models and to promote transparency in AI development for neurological applications.

Ethics

Acknowledgements

We acknowledge the contributions of the 24 expert annotators from 18 institutions who labeled the EEG recordings, as well as the Human Epilepsy Project (HEP) experts for gathering and scoring the HEP dataset. We thank the patients whose data contributed to this research. This work was supported by grants from the NIH (R01NS102190, R01NS102574, R01NS107291, RF1AG064312, RF1NS120947, R01AG073410, R01HL161253, R01NS126282, R01AG073598) and NSF (2014431), with additional funding for individual investigators as detailed in the full manuscript.

Conflicts of Interest

M.B.W. is a co-founder of and holds equity in Beacon Biosignals. Beacon Biosignals did not contribute funding nor played any role in the study. Dr. Benbadis is a medical director and advisor for Stratus EEG, an advisor for Synergy EEG, and a task force member on ambulatory EEG accreditation for ABRET. Dr. Karakis is a consultant for Ceribell, Epitel, UCB and GSK. Dr. Haider receives author royalties from UpToDate Inc, Spring Publishing. Dr. Jing receives author royalty from Springer Publishing.