Single-Trial EEG Classification: Decoding Timbre Perception

This repository contains scripts for classifying EEG data on a single-trial basis to decode timbre perception. The pipeline is designed for use with 3D preprocessed EEG data (trials × channels × time), stored in MATLAB .mat format.

🧠 Objective

To investigate whether timbre-related auditory information can be decoded from early EEG responses using supervised machine learning techniques.

---

🗃️ classification_Groupcv.py

• Loads and prepares EEG data for classification.
• Labels each trial based on stimulus condition (based on information already stored in .mat data matrix)
• Uses GroupKFold to split trials into training/testing sets while ensuring block-wise trial grouping (trials were presented in blocks of 5 per participant).
• Includes:
  • PCA-based dimensionality reduction
  • Multiple classifiers: LDA, SVM, k-NN, Gradient Boosting
  • GridSearchCV for hyperparameter tuning

---

🧪 feature_extraction.py

Includes several domain-informed EEG feature engineering methods:

1. compute_psd: Power spectral density across canonical EEG bands
2. erp_features: Peak/latency extraction from time-domain ERPs
3. compute_periodicity: Harmonic frequency power (e.g., 55 Hz, 110 Hz)
4. offsets_features: Post-stimulus slope and mean amplitude
5. peak_power:

---

📦 Data Format

• Input: 3D NumPy arrays or MATLAB .mat files
  Format: (n_trials, n_channels, n_timepoints)