Source-linked AI summary
The Temple University Hospital Seizure Detection Corpus
Vinit Shah, Eva von Weltin, Silvia Lopez, James Riley McHugh, Lily Veloso, Meysam Golmohammadi, Iyad Obeid, Joseph Picone
TL;DR
Clinical seizure-detection research lacks large, representative annotated EEG resources. The paper develops TUSZ by triaging a large EEG corpus and manually annotating selected recordings, producing a corpus with broad seizure morphology, multiple annotation views, and public access. TUSZ contains over 504 hours of data, including about 36 seizure hours, while seizure yield remains low in the source corpus.
Problem
Large annotated EEG resources are needed for deep-learning seizure detection, but existing resources are limited and high-performing results on small corpora have not represented clinical performance.
Method
The authors triaged TUH EEG sessions using report-based NLP and two automated seizure-detection systems, then manually annotated selected high-yield sessions.
Results
TUSZ contains over 504 hours of data, including about 36 hours of seizure events, and provides channel-based and term-based annotations across 315 subjects and 822 sessions.
Takeaways & Limitations
TUSZ combines varied seizure morphologies, multiple annotation views, and metadata in a publicly available corpus intended to support machine-learning research on clinical EEG data.
Abstract
from arXiv · showhide
We introduce the TUH EEG Seizure Corpus (TUSZ), which is the largest open source corpus of its type, and represents an accurate characterization of clinical conditions. In this paper, we describe the techniques used to develop TUSZ, evaluate their effectiveness, and present some descriptive statistics on the resulting corpus.
Method
TUSZ was built by selecting likely seizure-containing sessions from a large EEG corpus, then prioritizing and manually annotating high-yield data. The triage combined report-based NLP, commercial detection software, and AutoEEG.
- Data source: The source data comprised approximately 90% of TUH EEG v0.6.0, organized by patient and session with EDF recordings and neurologist reports.Annotation used 19 EEG channels plus heart rate and photic stimulation channels.
- Triage: Three independent approaches identified likely seizure sessions: NLP keyword search, Persyst P13, and the experimental AutoEEG system.Reports were normalized, stemmed, and classified using machine-learning and rule-based methods.
- Triage: NegEx correctly classified 99% of reports in a 100-report pilot, and flagged 844 likely seizure-containing sessions among 18,000 TUH EEG sessions.The rule-based approach modeled negation, context, and affirmation.
- Triage: Persyst identified seizure events in 1,388 of 34,698 files, while AutoEEG detected high-confidence seizures in 1,466 of 31,645 files.Files flagged by both systems received the highest annotation priority.
- Annotation: The triage selected 872 sessions containing 2,582 files from 16,168 sessions for manual annotation, yielding 280 sessions with actual seizures.Keyword search was the most effective triage approach; automated detectors performed poorly especially on short seizures.
Results
The resulting corpus provides annotated seizure recordings with multiple annotation views, seizure-type labels, metadata, and training/evaluation partitions. Its evaluation set was designed to represent training conditions while remaining compact for rapid experimentation.
- Corpus contents: TUSZ v1.2.0 contains 315 subjects and 822 sessions, including 280 sessions with seizures.Each file is transcribed using both channel-based and term-based annotations.
- Annotation views: Channel-based annotations mark event start and end times on specific channels, whereas term-based annotations aggregate annotations across channels.The term-based representation summarizes the channel-based labels into a shared annotation.
- Seizure labels: Seizure types were identified from neurologist reports and signal examination, with labels including generalized, focal, tonic-clonic, absence, and myoclonic seizures.When evidence was insufficient, events were labeled generalized non-specific or focal non-specific according to focality.
- Dataset partitioning: The database was divided into training and evaluation sets containing 265 and 50 subjects, respectively.The evaluation set was selected to represent training conditions while maximizing demographic features.
- Metadata and access: TUSZ includes raw EEG, seizure annotations, patient demographics, seizure type, EEG study type, and spreadsheet and time-aligned label-file resources.The label files support visualization of seizure labels with an open-source annotation tool.
Discussion
TUSZ provides a large, publicly available seizure-detection corpus with diverse seizure morphologies and multiple annotation views. Its construction and ongoing development address the challenges of representing clinical seizure data at scale.
- Over 504 hours of annotated EEG data are publicly available in TUSZ, including about 36 hours of seizure events, or 7% of the annotated data.
- The estimated seizure-data yield from the described process is 0.2%, making accurate triage important for cost-effective corpus construction.The authors note that the yield from continuous data is even smaller.
- TUSZ includes varied seizure morphologies across onset, termination, frequency, amplitude, locality, and focality.The authors state that this mix protects evaluation and training sets against bias toward one morphology.
- Event-based annotations capture per-channel seizure origin, spread, and termination, while term-based annotations record the earliest start and latest end times across channels.
- Multi-class annotations specify seizure type, whereas bi-class annotations indicate only whether a seizure is occurring.
- TUSZ has been downloadable in beta release since late 2016 and had over 700 registered users at the time reported.
Funding
The reported work received support from the National Human Genome Research Institute of the National Institutes of Health and the National Science Foundation.
- The publication was supported by the National Human Genome Research Institute of the National Institutes of Health under award U01HG008468.
- The work was also partly supported by the National Science Foundation under Grant No. IIP-1622765.
- The funding disclaimer states that the content and conclusions are the authors’ responsibility and may not represent the official views of the supporting agencies.