Single-Channel SSVEP BCI with Phase-Enhanced FBCCA

Overview

This repository contains the code and results for a study investigating whether explicit phase information from joint frequency-phase modulated (JFPM) stimuli can improve single-channel SSVEP-based brain-computer interface (BCI) decoding.

Key finding: Adding phase features to Filter Bank Canonical Correlation Analysis (FBCCA) improves single-channel classification accuracy from 62.30% to 65.38% at the Oz electrode (2-second window), with the largest gains (+8 percentage points) at shorter 1-second windows.

Authors

• Leonel Vázquez Carrasco
• Julie Verne Henriksen

Course: 22053 Principles of Brain-Computer Interfaces

Repository Structure

├── data/ # Dataset instructions (download manually)
| └── README.md # How to obtain the Benchmark SSVEP dataset
├── figures/ # Generated plots (SVG format)
├── results/ # CSV output files
│ ├── summary/ # Main results tables
│ ├── single_channel/ # Oz channel per-subject results
│ └── nine_channel/ # 9-channel baseline results
├── src/ # Python source code
│ └── legacy_original_script.ipynb # Runs all code
├── requirements.txt # Python dependencies
├── report.pdf # Full project report
└── README.md # This file

Requirements

Python environment

pip install -r requirements.txt

Dataset

This project uses the Benchmark SSVEP Dataset by Wang et al. (2017). The dataset is not included in this repository due to size and licensing restrictions.

To obtain the data:

1. Read the instructions in data/README.md
2. Download from the official source: http://bci.med.tsinghua.edu.cn/download.html
3. Place the .mat files in the data/ folder

Usage

Run FBCCA and FBCCA+Phase on all 9 occipital-parietal channels, the multi-channel FBCCA baseline and the time-window grid search:

python src/legacy_original_script.ipynb

Key Results

Single-channel performance (2-second window, Oz electrode)

Method	Accuracy (%)	ITR (bits/min)
FBCCA alone	62.30 ± 26.01	59.67 ± 33.51
FBCCA + Phase	65.38 ± 26.56	64.43 ± 34.74
Improvement	+3.08 (p < 0.0001)	+4.76 (p < 0.0001)

Time-window analysis (Oz electrode)

Window (s)	FBCCA (%)	FBCCA+Phase (%)	Δ (%)	p-value
0.75	20.48	27.71	+7.24	< 0.0001
1.00	31.56	39.56	+8.00	< 0.0001
1.50	49.45	54.82	+5.37	< 0.0001
2.00	62.30	65.55	+3.08	< 0.0001

9-channel baseline (2-second window)

Method	Accuracy (%)	ITR (bits/min)
FBCCA alone	90.50 ± 12.94	101.84 ± 22.45
FBCCA + Phase	90.38 ± 13.22	101.72 ± 22.91
Improvement	-0.12 (n.s.)	-0.12 (n.s.)

Phase features provide meaningful gains only in the single-channel (resource-constrained) setting.

Figures

All figures are available in figures/ as SVG files:

Figure	Description
Results_schannel.svg	Accuracy/ITR across 9 channels
Results_subject_sch.svg	Per-subject results for Oz channel
Results_subjects_9_channels.svg	Per-subject results for 9 channels
Results_subject_best_channel_per_subject.svg	Per-subject results for best individual channel
Results_tw_grid_All9.svg	Time-window grid search for 9 channels
Results_tw_grid_Oz.svg	Time-window grid search for single channel

Reproducing Results from Scratch

1. Clone this repository
2. Download the dataset (see data/README.md)
3. Install dependencies: pip install -r requirements.txt
4. Run the three main scripts (see Usage section above)
5. Generated results will match the CSV files in results/

License

This project is licensed under the MIT License - see the LICENSE file for details.

The dataset is provided for academic research purposes only; please cite the original authors.

Acknowledgments

• Professor Sadasivan Puthusserypady for guidance (DTU course 22053)
• Chen et al. (2015) for the original FBCCA implementation
• Wang et al. (2017) for the benchmark dataset
• All 35 subjects who participated in the original data collection