Notice: The reproducibility variables underlying each score are classified using an automated LLM-based pipeline, validated against a manually labeled dataset. LLM-based classification introduces uncertainty and potential bias; scores should be interpreted as estimates. Full accuracy metrics and methodology are described in Coakley et alK. L. Coakley, T. Snelleman, H. Hoos, and O. E. Gundersen, "The embrace of open science: An analysis of a decade of AI research and 56 800 conference papers," Under Review, 2026..
EvoBrain: Dynamic Multi-Channel EEG Graph Modeling for Time-Evolving Brain Networks
Authors: Rikuto Kotoge, Zheng Chen, Tasuku Kimura, Yasuko Matsubara, Takufumi Yanagisawa, Haruhiko Kishima, Yasushi Sakurai
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | The paper includes a dedicated section '5 Experiments and Results' which details experimental setup, metrics (AUROC, F1 score), baselines, and various performance tables and figures (Table 1, Figure 2, Figure 3, Figure 4, Figure 5) to evaluate the proposed Evo Brain model on seizure detection and prediction tasks. |
| Researcher Affiliation | Academia | All authors are affiliated with academic institutions: 'SANKEN, The Univerity of Osaka, Japan', 'Department of Neurosurgery, Graduate School of Medicine, The University of Osaka, Japan', and 'Institute for Advanced Co-Creation Studies, The University of Osaka, Japan'. The email domains also correspond to academic institutions: '@sanken.osaka-u.ac.jp' and '@nsurg.med.osaka-u.ac.jp'. |
| Pseudocode | No | The paper describes the proposed method, Evo Brain, in Section 4, detailing its components like Temporal Modeling with Mamba and Spatial Modeling with GCN and Lap PE, using mathematical formulations and textual descriptions. However, it does not include any explicitly labeled pseudocode or algorithm blocks. |
| Open Source Code | Yes | Under 'E Experimental details and implementation', the authors state: 'We release Git Hub repository (https://github.com/Kotoge/Evo Brain).' |
| Open Datasets | Yes | The paper states under '5.1 Experimental Setup': 'Datasets. We used the Temple University Hospital EEG Seizure dataset v1.5.2 (TUSZ) (Shah et al., 2018) to evaluate Evo Brain. ... Additionally, we used the smaller CHB-MIT dataset, which consists of 844 hours of 22-channel scalp EEG data from 22 patients, including 163 recorded seizure episodes.' Both TUSZ and CHB-MIT are well-known public datasets. |
| Dataset Splits | Yes | Under '5.1 Experimental Setup', the paper specifies: 'For the TUSZ dataset, we followed the official data split, in which a subset of patients is designated for testing. Similarly, for the CHB-MIT dataset, we used randomly selected 15% of the patient s data for testing.' Additionally, 'Table 3: Number of EEG data samples and patients in the train, validation, and test sets on TUSZ dataset. Train, validation, and test sets consist of distinct patients.' provides detailed numerical splits for training, validation, and test sets. |
| Hardware Specification | Yes | Under 'E Experimental details and implementation', the paper explicitly states: 'Training for all models was accomplished using the Adam optimizer (Kingma and Ba, 2014) in Py Torch on NVIDIA A6000 GPU and Xeon Gold 6258R CPU.' |
| Software Dependencies | No | The paper mentions 'Py Torch' in the 'E Experimental details and implementation' section: 'Training for all models was accomplished using the Adam optimizer (Kingma and Ba, 2014) in Py Torch on NVIDIA A6000 GPU and Xeon Gold 6258R CPU.' However, it does not provide a specific version number for PyTorch or any other key software libraries, which is necessary for a reproducible description. |
| Experiment Setup | Yes | Under 'E Experimental details and implementation', the 'Implementation details' section provides specific hyperparameters and training configurations: 'We used binary cross-entropy as the loss function to train all models. The models were trained for 100 epochs with an initial learning rate of 1e-4. To enhance efficiency and sparsity, we set τ = 3 and the top-3 neighbors edges were kept for each node. The dropout probability was 0 (i.e., no dropout). Evo Brain has two Mamba consisting of two stacked layers and a two-layer GCN with 64 hidden units, resulting in 114,794 trainable parameters.' |