Pre-Training Graph Contrastive Masked Autoencoders are Strong Distillers for EEG

Authors: Xinxu Wei, Kanhao Zhao, Yong Jiao, Hua Xie, Lifang He, Yu Zhang

ICML 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We validate the effectiveness of EEG-Dis GCMAE across four classification tasks using two clinical EEG datasets with abundant data. ... We evaluated our EEG-Dis GCMAE framework on two clinical datasets with rs-EEG time series: the Establishing Moderators and Biosignatures of Antidepressant Response in Clinical Care (EMBARC) (Trivedi et al., 2016) and the Healthy Brain Network (HBN) (Alexander et al., 2017). ... 5. Experiments ... 5.4. Ablation Study Analysis ... 6. Experiments for Model Robust Analysis
Researcher Affiliation Academia 1Department of Electrical and Computer Engineering, Lehigh University, Bethlehem, PA, USA 2Department of Bioengineering, Lehigh University, Bethlehem, PA, USA 3Center for Neuroscience Research, Children s National Hospital, Washington, DC, USA 4Department of Computer Science and Engineering, Lehigh University, Bethlehem, PA, USA.
Pseudocode Yes Algorithm 1 GTD Loss Calculation Input: X l, Vl, Al, X h, Vh, Ah, Vd Parameter: F(.), θ, ϵ Output: LGT D Dis 1: Normalize Al, Ah 2: Apply threshold: Al (Al > θ), Ah (Ah > θ) 3: Compute kernel matrices: Zl = F(X l), Zh = F(X h) 4: Assert: |Vl| |Vh| 5: if |Vl| = |Vh| then 6: Extract sub-matrices Ah sub = Ah[Vl, Vl] 7: Direct connections: Ah 1 hop = (Ah sub > 0) 8: Indirect connections: Ah 2 hop = Ah[Vd, : Vl] 9: LP os = KL(Zl||Zh) | (Ah 1 hop Ah 2 hop) 10: LNeg = KL(Zl||Zh) | (Al (Ah 1 hop Ah 2 hop)) 11: else 12: LP os = KL(Zl||Zh) | (Ah > 0) 13: LNeg = KL(Zl||Zh) | (Al > 0 Ah = 0) 14: end if 15: LP os Avg = LP os Cpos 16: LNeg Avg = LNeg Cneg 17: return LGT D Dis = LP os Avg LNeg Avg+ϵ
Open Source Code Yes The source code is available at https://github.com/ weixinxu666/EEG_Dis GCMAE.
Open Datasets Yes We evaluated our EEG-Dis GCMAE framework on two clinical datasets with rs-EEG time series: the Establishing Moderators and Biosignatures of Antidepressant Response in Clinical Care (EMBARC) (Trivedi et al., 2016) and the Healthy Brain Network (HBN) (Alexander et al., 2017).
Dataset Splits Yes For downstream task fine-tuning, due to the limited amount of labeled data, we employed 10-fold cross-validation with 10 runs for all model training.
Hardware Specification No Portions of this research were conducted on Lehigh University s Research Computing infrastructure partially supported by NSF Award 2019035.
Software Dependencies No Both pre-training and fine-tuning were optimized using the Adam optimizer.
Experiment Setup Yes During pre-training, we used a batch size of 128. For downstream fine-tuning, we used a batch size of 32. Both pretraining and fine-tuning were optimized using the Adam optimizer. ... Pre-training was performed over 200 epochs, while downstream fine-tuning was carried out for 400 epochs.