NEO-8D-code

Analysis code for the NEO-8D project: decoding single and dual upper-limb movements from ECoG spectral features.

Two participants (P01, P02) performed four single movements (Left/Right Elbow, Left/Right Hand) and four dual movements (combinations of two single movements). This repository contains the high-gamma PSD datasets together with self-contained MATLAB scripts covering the main analyses: ERDS spectra, demixed PCA, linear decomposition of dual-movement patterns, class-distance clustering, cross-validated decoding, and an offline center-out BCI simulation.

Requirements

• MATLAB R2023b (older releases down to ~R2019b should also work).
• Statistics and Machine Learning Toolbox (pca, fitlm, fitcdiscr, linkage, ksdensity, ...).
• No third-party toolboxes: all colormaps are vendored in utils/slanCM_palette.m.

Data download

Please download the two PSD datasets (PSDdata-P01.mat, PSDdata-P02.mat) from the figshare link and place them in the data/ folder:

NEO-8D-code/
└── data/
    ├── PSDdata-P01.mat
    └── PSDdata-P02.mat

Each file contains three variables:

Variable	Size	Description
Spectra	freq x channels x trials	PSD values per trial (8 ECoG channels)
fb	76 x 1	frequency axis, ~2–148 Hz (0.5 s window, fs = 1000 Hz)
Ylab	trials x 1	trial label (see table below)

Trial labels:

Code	Condition	Code	Condition
100	Rest	121	LE & LH
101	LE (left elbow)	122	LE & RE
102	LH (left hand)	125	LH & RH
103	RE (right elbow)	126	RE & RH
104	RH (right hand)

• P01: 5588 trials (motor-imagery style cues)
• P02: 1912 trials (attempted movements)

How to run

1. Open MATLAB in the NEO-8D-code/ folder (startup.m adds utils/ to the path automatically; if you start MATLAB elsewhere, run startup once first).

2. Every script is standalone — open it and run top-to-bottom. Set the participant at the top of the file:

   SUBJECT = 'P01';   % 'P01' | 'P02'

3. Or run a script for a specific participant without editing it:

   SUBJECT = 'P02';  offline_bci_simulation

Scripts

Script	What it does	Approx. runtime
erds_spectra.m	Task-related SNR (ERDS) spectra per movement class for both participants; optionally HG RMS time courses if non-public HGseqdata-*.mat files are placed in data/	< 1 min
movement_dpca.m	Demixed PCA on single-movement patterns; projects single and dual trials onto the first two demixed axes	< 1 min
dual_psd_decomposition.m	Regresses each dual-movement PSD pattern onto linear combinations of single-movement patterns (R² bar plot, regression-weight scatter plots, paired statistics)	< 1 min
class_distance_clustering.m	Pairwise Mahalanobis distances between the 8 movement classes + Ward hierarchical clustering	< 1 min
class_psd_maps.m	Trial-averaged channel x frequency PSD maps for each of the 8 classes	< 1 min
decoding_confusion_matrices.m	Repeated 80/20 LDA decoding (100 splits): single-movement 4-class and merged single+dual 4-class confusion matrices; population-vector diagram	a few minutes
offline_bci_simulation.m	Full offline pipeline: 3 decoding schemas + stochastic center-out BCI simulation driven by the measured confusion matrices (5040 population-vector permutations)	~10–30 min

Notes

• Axis limits in some scripts are tuned for P01; for P02 the scripts fall back to automatic limits where relevant.

Reference

(add the manuscript citation here once available)