Schizsgorithm: EEG-Based Schizophrenia Classification

Schizsgorithm is a machine learning pipeline for classifying schizophrenia from EEG recordings. The project uses MNE for EEG preprocessing, extracts statistical features from fixed-length signal windows, and evaluates classification performance with subject-aware cross-validation to reduce leakage across individuals.

---

Overview

EEG is a noisy, high-variance biosignal with substantial variability across subjects and recording sessions. This project explores whether schizophrenia-related patterns can be identified from EEG recordings using a reproducible preprocessing and feature-based classification pipeline.

The core idea is:

1. Load raw EEG recordings from EDF files
2. Preprocess the signal using MNE
3. Split continuous recordings into fixed-length epochs
4. Extract statistical features from each epoch
5. Train and evaluate a classifier using group-aware cross-validation

---

Pipeline

1. Data Loading

EEG recordings are loaded from EDF files using mne.io.read_raw_edf.

• Files beginning with h are treated as healthy controls
• Files beginning with s are treated as schizophrenia cases

2. Preprocessing

Each EEG recording is preprocessed with:

• EEG re-referencing
• Band-pass filtering from 0.5 Hz to 60 Hz
• Segmentation into 5-second fixed-length epochs with 1-second overlap

This converts each subject recording into multiple fixed-length EEG windows.

3. Feature Extraction

For each epoch, the pipeline computes statistical features channel-wise, including:

• Mean
• Standard deviation
• Peak-to-peak amplitude
• Variance
• Minimum
• Maximum
• Argmin
• Argmax
• RMS
• Absolute difference signal
• Skewness
• Kurtosis

These features are concatenated into a single feature vector for classification.

4. Modeling

The current implementation uses:

• StandardScaler
• LogisticRegression
• GridSearchCV for tuning the regularization parameter C

5. Evaluation

The model is evaluated using GroupKFold, where each subject is treated as a separate group.

This is important because it prevents epochs from the same subject from appearing in both training and validation folds, reducing subject leakage and giving a more realistic estimate of generalization.

---

Why Group-Aware Validation Matters

A major risk in EEG classification is overly optimistic performance caused by leakage across windows from the same subject. This project explicitly addresses that by grouping all epochs from a single subject together during cross-validation.

That means the classifier is evaluated more like a real-world system:

• train on some subjects
• test on unseen subjects

---

Current Method

Preprocessing

• Re-reference EEG
• Filter from 0.5 to 60 Hz
• Epoch into 5-second windows with 1-second overlap

Features

Handcrafted statistical descriptors computed per channel and concatenated into one feature vector.

Classifier

• Logistic Regression
• Hyperparameter tuning over:
  • C = [0.1, 0.6, 2, 3, 4, 7, 34]

Cross-Validation

• GroupKFold(n_splits=5)

---

Repository Structure

.
├── datafiles/              # EDF files
├── main.py                 # training / evaluation pipeline
└── README.md