EEG Person Identification via Hybrid CNN-BiLSTM

This repository contains a deep learning pipeline for identifying individuals based on their EEG (Electroencephalography) brainwave patterns. It utilizes a hybrid architecture combining Convolutional Neural Networks (CNNs) for spatial/frequency feature extraction and Bidirectional LSTMs for temporal dynamics.

Project Overview

Biometric identification using EEG signals offers a high-security alternative to traditional methods as brainwaves are difficult to spoof. This project implements a robust pipeline that transforms raw EEG signals into Time-Frequency spectrograms and classifies them using a deep neural network.

Dataset

This project uses the EEG Motor Movement/Imagery Dataset provided by PhysioNet.

• Source: PhysioNet EEG Motor Movement/Imagery Dataset
• Subjects: 109 volunteers.
• Structure: 64-channel EEG recordings sampled at 160 Hz.
• Split:
  • Training: Session A (Runs 1–7)
  • Testing: Session B (Runs 8–14)

The Pipeline

1. Preprocessing (MNE & SciPy)

Raw EDF files are processed to remove noise and standardize inputs:

• Filtering: FIR Bandpass filter (1–40 Hz) to isolate relevant brain rhythms (Delta through Beta).
• Epoching: Signals are segmented into 2.0-second windows with a 0.5-second step.
• Feature Extraction: Short-Time Fourier Transform (STFT) is applied to generate spectrograms, converting 1D signals into 2D images (Frequency × Time).

2. Model Architecture

The model (CNNBiLSTM) is designed to capture spatiotemporal features:

1. CNN Encoder:
   • Stack of 2D Convolutional layers with Batch Normalization and ReLU.
   • Asymmetric Pooling: Max pooling is applied primarily to the Frequency dimension, preserving the Time dimension to allow the LSTM to analyze temporal changes.
2. Temporal Processing:
   • A Bidirectional LSTM processes the sequence of feature vectors extracted by the CNN.
3. Classification:
   • A dense classification head maps the final LSTM states to subject identities.

Requirements

The project relies on the following major libraries:

• Python 3.11+
• PyTorch (Model training)
• MNE (EEG signal loading and filtering)
• NumPy & Pandas (Data manipulation)
• SciPy (Signal processing/STFT)
• Scikit-Learn (Metrics and Label Encoding)

Usage

The pipeline is contained within eeg-person-id.ipynb.

1. Configure Paths: Update the Config class in the notebook to point to your local dataset path.

   Data_Root = Path("/path/to/eeg-motor-movementimagery-dataset")

2. Run Pipeline: Set RUN_PIPELINE = True to process raw data into .npy arrays.
3. Train: Run the training cells to train the CNN-BiLSTM model.
4. Evaluate: The notebook automatically generates a classification report, confusion matrix, and t-SNE plot.

Results

• Visualization: The notebook includes t-SNE visualization to project the high-dimensional latent embeddings into 2D, demonstrating how the model clusters signals by subject identity.
• Metrics: Standard accuracy, Precision, Recall, and F1-Score are calculated per subject.