Brain Computer Interface AI Model

AI model for a Brain Computer Interface. Detects left/right imagined movement from 3 EEG electrodes (C3, C4, Cz). Applies a convolutional network with temporal filters, then spatial filters (across electrodes), followed by another convolutional layer, and dense layer. Validation accuracy ~76%.

Based on approach in https://arxiv.org/pdf/1611.08024.
Training data: https://www.bbci.de/competition/iv/desc_2b.pdf

Package install

pip install -r requirements.txt

Inference

EEG data should be a numpy array of size [3, 384] corresponding to channels (C3, C4, Cz) x time series (3s of data sampled at 128 Hz, bandpass 4-50 Hz, in volts). EEG data epoch should start sync'd to stimulus trigger (tone 1 kHz, 70 ms plus prompt, e.g. "Think left or right")

from inference import EEGInferenceApp

app = EEGInferenceApp("data/checkpoint.pt")
movement = app.predict_imagined_movement(data)

if movement == EEGInferenceApp.LEFT:
    print("You imagined left")
elif movement == EEGInferenceApp.RIGHT:
    print("You imagined right")

Training

python training.py

Data preprocessed into data/preprocessed_data.pt
Checkpoint saved into data/checkpoint.pt