🧠 High-gamma and Low-frequency ECoG signal Fusion Network (HiLoFuseNet) for continuous finger movement decoding

This repository provides the official PyTorch implementation of the finger movement decoding framework detailed in the paper:

Sun et al., "Spectro-Temporal Fusion of High-Gamma and Low-Frequency ECoG Signals for Intracranial Finger Movement Decoding," 2025. under review.

The model achieved SOTA decoding performance on the public BCI Competition IV and Stanford datasets, with Pearson correlation coefficients between true and predicted finger movement trajectories of 0.631 and 0.534, representing improvements of 5.0% and 11.9%, respectively, over the previous best methods.

📢📢📢 News

• Our original submission to TNNLS was rejected. The reviewers raised concerns regarding the state-of-the-art claim and the necessity of fusing low-frequency signals. We are currently improving the decoding framework and conducting thorough comparisons against the latest decoders.

⚠️⚠️⚠️ We need you!

We benchmarked a large number of decoders across the two datasets. Since most of them did not open-source their code, despite our best efforts to replicate the reported results, the discrepancies we observed were substantial. If, by any chance, you find that our implementation is incorrect, please email us!

---

🛠️ Decoding Framework

The proposed framework is characterized by (a) a streamlined ECoG feature extraction pipeline and (b) a compact neural network for learning spectro-temporal information.

Core Functions

Functionality	Implementation Path
HGA and LFS Feature Extraction	finger_regression/models/prepareDataset.py/HGALFS_feature_extractor
HiLoFuseNet Architecture	finger_regression/models/nn_regressors.py/HiLoFuseNet

---

💾 Quick Start & Reproducibility

1. Download Datasets

• BCIIV: https://www.bbci.de/competition/iv/#dataset4
• Stanford-FingerFlex: https://searchworks.stanford.edu/view/zk881ps0522

2. Signal Preprocessing

Raw ECoG signals were preprocessed using MATLAB FieldTrip-20230926.

• BCIIV Preprocessing: data_preprocessing/data_preprocessing_BCI4.m
• Stanford Preprocessing: data_preprocessing/data_preprocessing_Stanford.m

3. Feature Extraction

Features were extracted using MNE-Python (v1.8.0).

• Script: finger_regression/prepare_taskFormatedData.py

4. Run Experiments

Configure the pytorch environment via finger_regression/environment.yml. The following table summarizes the scripts used to reproduce the paper's findings. Raw output files are provided in finger_regression/results. The .slum file provides the code to interact with a supercomputing cluster. If you run the script locally, please change the inputs in the .py file according to the corresponding settings from the .slurm file.

Experiment	Execution Script(s)	Results Folder
DNN Multi-Output Regression	regression_o5_nn.py, submit_o5_nn.slurm	finger_regression/results/o5/varyingSeed
ML Multi-Output Regression	regression_o5_ml.py, submit_o5_ml.slurm	finger_regression/results/o5
Model Interpretation	regression_o5_nn_interpretModel.py, submit_o5_nn_interpretModel.slurm	finger_regression/results/o5/interpretModel
Ablation Study	regression_o5_nn_ablation.py, submit_o5_nn_ablation.slurm	finger_regression/results/o5/ablation
Hyperparameter Test	regression_o5_nn_hyperparameter.py, submit_o5_nn_hyperparameter.slurm	finger_regression/results/o5/hyperparameter
DNN Single-Output Regression	regression_o1_nn.py, submit_o1_nn.slurm	finger_regression/results/o1/varyingSeed

Acknowledgement

A sincere thanks to the code contributors for BTTR and HOPLS!

• BTTR: https://github.com/TheAxeC/block-term-tensor-regression
• HOPLS: https://github.com/arthurdehgan/HOPLS

Citation

Hope this model helps your research. We would be appreciated if u cite us.

to be appeared...