66 lines
2.8 KiB
Markdown
66 lines
2.8 KiB
Markdown
# Reproduction: Mwata-Velu et al. (2023)
|
||
|
||
**Paper**: "EEG-BCI Features Discrimination between Executed and Imagined Movements
|
||
Based on FastICA, Hjorth Parameters, and SVM"
|
||
**Journal**: Mathematics 2023, 11, 4409
|
||
**DOI**: https://doi.org/10.3390/math11214409
|
||
|
||
## Overview
|
||
|
||
This repository contains an attempted reproduction of the above paper as part of a
|
||
thesis on reproducibility challenges in EEG-based BCI research. The reproduction was
|
||
**partially completed** — the core pipeline is implemented but several ambiguities in the
|
||
paper prevented a definitive reproduction.
|
||
|
||
## Repository Structure
|
||
|
||
```
|
||
config.py — Constants, channel lists, run mappings, parameters
|
||
data_loading.py — CSV data loading, MNE Raw creation, annotations
|
||
pipeline.py — Filtering, FastICA, Hjorth features, SVM classification
|
||
reproduction_notebook.ipynb — Main analysis notebook (Method 2: cross-subject)
|
||
requirements.txt — Python dependencies
|
||
```
|
||
|
||
## Data
|
||
|
||
This code expects the PhysioNet EEG Motor Movement/Imagery Dataset in the curated
|
||
CSV format provided by:
|
||
Z. Shuqfa, A. Lakas, and A. N. Belkacem, “Increasing accessibility to a large brain–
|
||
computer interface dataset: Curation of physionet EEG motor movement/imagery dataset
|
||
for decoding and classification,” Data in Brief, vol. 54, p. 110181, Jun. 2024,
|
||
doi: 10.1016/j.dib.2024.110181.
|
||
|
||
Files are named:
|
||
- `eegmmidb/SUB_001_SIG_01.csv` — Signal data (n_samples × 64 channels)
|
||
- `eegmmidb/SUB_001_ANN_01.csv` — Annotations (label, duration, start/end rows)
|
||
|
||
The curated dataset excludes the 6 problematic subjects (S088, S089, S092, S100,
|
||
S104, S106) noted in the paper. Run numbering is offset by 2 from PhysioNet's
|
||
original (our Run 01 = PhysioNet R03).
|
||
|
||
## Usage
|
||
|
||
1. Install dependencies: `pip install -r requirements.txt`
|
||
2. Place curated CSV data in `eegmmidb/` directory
|
||
3. Edit `config.py` to set `ICA_STRATEGY` ('per_run', 'per_subject', or 'global')
|
||
4. Run `reproduction_notebook.ipynb`
|
||
|
||
## Key Implementation Decisions
|
||
|
||
| Decision | Paper says | We do | Rationale |
|
||
|----------|-----------|-------|----------|
|
||
| Pipeline order | Figure 1: filter→ICA; Algorithm 1: ICA with internal sub-band eval | ICA then sub-band eval | Energy criterion is meaningless on pre-filtered data |
|
||
| ICA algorithm | Gram-Schmidt (Algorithm 1, Step 3) | `algorithm='deflation'` | Deflation uses Gram-Schmidt |
|
||
| Energy criterion | ∀χ ∈ {α, β, **γ**} | ∀χ ∈ {**θ**, α, β} | γ never defined; likely typo for θ |
|
||
| ICA scope | Not specified | Configurable | Reproducibility variable |
|
||
| Classification Method | Methods 1 and 2 | Method 2 only (cross-subject) | Method 1 split is contradictory |
|
||
|
||
## Paper's Reported Results (Method 2, Set 3)
|
||
|
||
| Metric | Paper |
|
||
|--------|-------|
|
||
| Overall accuracy | 68.8 ± 0.71% |
|
||
| ME recall | 68.17% |
|
||
| MI recall | 68.41% |
|