Self-Supervised Graph Neural Networks for Improved Electroencephalographic Seizure Analysis
Authors: Siyi Tang, Jared Dunnmon, Khaled Kamal Saab, Xuan Zhang, Qianying Huang, Florian Dubost, Daniel Rubin, Christopher Lee-Messer
ICLR 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | When evaluating our approach on seizure detection and classification on a large public dataset (5,499 EEGs), we find that our GNN with self-supervised pre-training achieves 0.875 Area Under the Receiver Operating Characteristic Curve on seizure detection and 0.749 weighted F1-score on seizure classification, outperforming previous methods for both seizure detection and classification. |
| Researcher Affiliation | Academia | Siyi Tang, Jared Dunnmon, Khaled Saab, Xuan Zhang , Qianying Huang , Florian Dubost, Daniel Rubin , Christopher Lee-Messer Stanford University, CA, USA {siyitang,jdunnmon,ksaab,kayleez,qyhuang,fdubost,rubin,cleemess} @stanford.edu |
| Pseudocode | No | The paper does not contain any clearly labeled pseudocode or algorithm blocks. |
| Open Source Code | Yes | Source code is publicly available at https://github.com/tsy935/eeg-gnn-ssl. |
| Open Datasets | Yes | We use the public Temple University Hospital EEG Seizure Corpus (TUSZ) v1.5.2 (Shah et al., 2018; Obeid & Picone, 2016), the largest public EEG seizure database to date with 5,612 EEGs, 3,050 annotated seizures from clinical recordings, and eight seizure types. ... The Temple University Hospital EEG Seizure Corpus used in our study is publicly available3 with full IRB approval (Shah et al., 2018; Obeid & Picone, 2016). (Footnote 3: https://www.isip.piconepress.com/projects/tuh_eeg/html/downloads.shtml) |
| Dataset Splits | Yes | We randomly split the official TUSZ train set by patients into train and validation sets by 90/10 for model training and hyperparameter tuning, respectively, and we hold-out the official TUSZ test set for model evaluation (excluding five patients who exist in both the official TUSZ train and test sets). The train, validation, and test sets consist of distinct patients. See Appendix D for the number of preprocessed EEG clips and patients in each split. |
| Hardware Specification | Yes | Training for all models was accomplished using the Adam optimizer (Kingma & Ba, 2014) in Py Torch on a single NVIDIA Titan RTX GPU. |
| Software Dependencies | No | The paper mentions 'Py Torch' and 'Sci Py python package' with citations (Virtanen et al., 2020a, 2020b) which refer to SciPy 1.0. However, a specific version number for PyTorch is not provided, and the prompt requires multiple key software components to have specific version numbers. |
| Experiment Setup | Yes | We performed the following hyperparameter search on the validation set: (a) initial learning rate within range [5e-5, 1e-3]; (b) τ {2, 3, 4}, the number of neighbors to keep for each node in the correlation graphs; (c) the number of Diffusion Convolutional Gated Recurrent Units (DCGRU) layers within range {2, 3, 4, 5} and hidden units within range {32, 64, 128}; (d) the maximum diffusion step K {2, 3, 4}; (e) dropout probability in the last fully connected layer. We used a batch size of 40 EEG clips... |