Skip to content

Event-Notes-Day4.md

Latest commit

 

History

History
118 lines (74 loc) · 3.57 KB

Event-Notes-Day4.md

File metadata and controls

118 lines (74 loc) · 3.57 KB
Raw

📝 Event Notes– Day 4

Day 4 – Wednesday, 11 September 2024

Machine Learning for EEG in BCIs and Healthcare

Session 4: Machine Learning in BCIs and Healthcare — Dr. Mohamed Abdelhack

Key Takeaways:

  • EEG → preprocessed → feature extraction → classifier → BCI output.

  • Applications of ML in BCIs:

    • SSVEP-based spellers
    • Motor imagery BCIs
    • Neurological disorder detection

  • Importance: Careful preprocessing and feature selection are critical for accurate classification.

  • Emphasis on understanding SSVEP response characteristics and how ML can interpret them for BCIs.

Tutorial 4: Post-processing Analysis of EEG with Machine Learning — Dr. Nour El-Madany

Step-by-Step Workflow:

1. Sampling & Filtering

  • Converts continuous EEG signals to discrete-time by sampling at defined intervals.
  • Band-pass filtering (6–80 Hz) removes unwanted frequency components.
  • Goal: reduce noise and retain relevant neural oscillations.

2. Epoch Segmentation

  • EEG divided into short epochs (time windows) per trial.
  • Allows computation of features for each epoch.

3. Feature Extraction

  • Power Spectrum Density (PSD): energy at different frequencies (key for SSVEP).
  • Time-domain features: amplitude, latency of ERP components.
  • Frequency-domain features: band-power, dominant frequencies.

4. Machine Learning Classification

  • ML models classify EEG epochs using extracted features.
  • Example: Canonical Correlation Analysis (CCA) compares EEG with reference stimuli to identify the attended frequency.
  • Other possible models: LDA, SVM, deep learning.
  • Purpose: predict user intent or detect brain states.

5. Evaluation

  • Compare predicted labels with ground truth.
  • Metrics: classification accuracy, confusion matrices.

Workshop 4: Post-processing Analysis — Dr. Mohamed Zaky & Eng. Mai Gamal

Focus: Analysis of pre-recorded SSVEP dataset, post-processing and ML classification.

1. Dataset Setup

  • 12 SSVEP stimulus classes
  • 8 EEG channels
  • 15 trials per stimulus

2. Preprocessing

  • Band-pass filtering (6–80 Hz)
  • Artifact removal (ICA or ASR)
  • Epoch segmentation (4-second windows)

3. Frequency-Domain Analysis

  • Compute magnitude spectrum per epoch using FFT
  • Inspect key channels (e.g., Oz) for stimulus frequency peaks
  • Identify clear vs noisy SSVEP responses

4. Feature Extraction & Classification

  • CCA Workflow:

    1. Generate sine/cosine templates for each stimulus frequency and harmonics
    2. Compute correlation between EEG epochs and templates
    3. Assign stimulus label with highest correlation

  • Complete ML pipeline:

    1. Load EEG epochs
    2. Apply preprocessing & filtering
    3. Generate reference templates
    4. Apply CCA classification
    5. Evaluate accuracy per subject

5. Results

  • Accuracy varied across subjects:

    • Low (~26–30%)
    • High (~91%)

  • Average accuracy: 60%

  • Highlights inter-subject variability and the importance of preprocessing.

6. Key Takeaways

  • Focused on post-processing EEG data and ML-based classification for SSVEP BCIs.
  • Hands-on experience covered the full pipeline: Filtering → Epoching → Feature Extraction → CCA Classification → Evaluation
  • Emphasis on practical application using real pre-recorded datasets.

Outcome: By the end of the day, participants gained practical skills in SSVEP data analysis, feature extraction, ML classification, and evaluation of EEG-based BCIs,