wavelet-scattering-schizophrenia-eeg-biomarker

name	wavelet-scattering-schizophrenia-eeg-biomarker
description	Wavelet Scattering Transform (WST) framework for interpretable schizophrenia biomarker discovery and classification from resting-state EEG. Multi-order scattering coefficients capture cross-frequency coupling and amplitude modulation dynamics, achieving 90.48% accuracy under strict subject-independent evaluation.
tags	["eeg","schizophrenia","biomarker","wavelet-scattering-transform","cross-frequency-coupling","interpretable-ml","clinical-neuroscience"]
arxiv_id	2607.05282
date	2026-07-06T00:00:00.000Z
authors	["Unknown"]

Wavelet Scattering Transform for Schizophrenia EEG Biomarker Discovery

Core Innovation

Problem: Existing EEG-based schizophrenia classifiers rely on static power spectral density (PSD) features that are blind to amplitude modulation dynamics and cross-frequency coupling—phenomena central to schizophrenia pathophysiology. They also suffer from temporal data leakage through epoch-level cross-validation.

Solution: Multi-order Wavelet Scattering Transform (WST) with strict Leave-One-Subject-Out (LOSO) cross-validation and SHAP explainability.

Key Findings

1. Second-order scattering coefficients (encoding cross-frequency coupling) dominate discriminative biomarkers
2. Gamma-band features most prevalent in the biomarker set
3. Electrode P3 identified as single most discriminative site
4. 90.48% accuracy (AUC = 0.9339, sensitivity = 95.56%) under rigorous subject-independent evaluation

Methodology

1. Wavelet Scattering Transform (WST)

The WST is a mathematically principled feature extraction method that captures multi-scale amplitude modulation structure:

Zeroth-order: Local average (low-pass)

S[0]x(t) = x * φ_J(t)

First-order: Modulus of wavelet coefficients

S[1]x(t) = |x * ψ_j| * φ_J(t)

Second-order: Modulus of first-order coefficients

S[2]x(t) = ||x * ψ_j| * ψ_k| * φ_J(t)

Key Properties:

• Translation invariance: Stable to time shifts
• Deformation stability: Robust to small temporal warping
• Low variance: Statistical stability for classification
• Interpretability: Each order captures specific temporal scales

2. Feature Extraction Pipeline

import numpy as np
from kymatio import Scattering1D

# Initialize WST
J = 6  # Number of scales
Q = 8  # Wavelets per octave
scattering = Scattering1D(J=J, shape=signal_length, Q=Q)

# Extract scattering coefficients
# Input: (batch, channels, time)
Sx = scattering(eeg_data)

# Sx contains:
# - S[0]: 1 coefficient per channel (local average)
# - S[1]: J*Q coefficients per channel (first-order moduli)
# - S[2]: J*(J-1)/2*Q^2 coefficients per channel (second-order)

3. Biomarker Discovery

Statistical Testing:

from scipy import stats
from statsmodels.stats.multitest import multipletests

# Subject-level ANOVA for each feature
f_stats, p_values = stats.f_oneway(
    patients_features, controls_features, axis=0
)

# FDR correction
reject, pvals_corrected, _, _ = multipletests(
    p_values, alpha=0.05, method='fdr_bh'
)

significant_biomarkers = np.where(reject)[0]

4. Classification with Strict LOSO

from sklearn.ensemble import RandomForestClassifier
from sklearn.model_selection import LeaveOneGroupOut

logo = LeaveOneGroupOut()
classifier = RandomForestClassifier(n_estimators=100, random_state=42)

# Subject-level majority voting
for train_idx, test_idx in logo.split(X, y, groups=subject_ids):
    X_train, X_test = X[train_idx], X[test_idx]
    y_train, y_test = y[train_idx], y[test_idx]
    
    classifier.fit(X_train, y_train)
    
    # Epoch-level predictions
    epoch_preds = classifier.predict_proba(X_test)
    
    # Subject-level majority vote
    subject_pred = np.argmax(np.mean(epoch_preds, axis=0))

5. SHAP Explainability

import shap

# Create explainer
explainer = shap.TreeExplainer(classifier)
shap_values = explainer.shap_values(X_test)

# Identify top biomarkers
top_features = np.argsort(np.abs(shap_values).mean(axis=0))[-10:]

Critical Methodological Points

1. Avoid Temporal Data Leakage

WRONG: Epoch-level cross-validation

# ❌ This leaks information
for train_idx, test_idx in KFold().split(X_epochs, y_epochs):
    ...

CORRECT: Subject-level cross-validation

# ✓ No leakage
for train_idx, test_idx in LeaveOneGroupOut().split(X, y, groups=subjects):
    ...

2. Subject-Level Majority Voting

When classifying multiple epochs per subject, aggregate predictions:

# Get epoch-level probabilities
epoch_probs = classifier.predict_proba(X_subject)

# Majority vote
subject_prediction = np.argmax(np.mean(epoch_probs, axis=0))

3. Interpretation of Scattering Orders

• S[0]: Slow trends, DC offset
• S[1]: Amplitude modulations at scale j
• S[2]: Cross-frequency coupling between scales j and k

Schizophrenia-specific: Second-order coefficients capture disrupted gamma-theta coupling.

Implementation Checklist

• Use strict LOSO cross-validation (not epoch-level)
• Extract multi-order scattering coefficients (at least order 2)
• Apply FDR correction for multiple comparisons
• Use subject-level majority voting for final predictions
• Report both accuracy and AUC
• Include SHAP analysis for biomarker interpretability
• Validate on independent dataset if possible

Performance Benchmarks

Method	Accuracy	AUC	Cross-Validation
PSD + SVM	~75%	~0.80	Epoch-level (leaked)
PSD + RF	~78%	~0.83	Epoch-level (leaked)
WST + RF	90.48%	0.9339	LOSO (strict)
WST + SVM	88.12%	0.9102	LOSO (strict)

Clinical Implications

1. Objective Biomarker: First rigorous EEG-based schizophrenia biomarker with subject-independent validation
2. Electrode P3: Single-electrode classification possible, enabling simplified clinical protocols
3. Cross-Frequency Coupling: Confirms gamma-theta coupling disruption as core pathophysiology
4. Interpretability: SHAP analysis reveals which temporal dynamics are most discriminative

Pitfalls and Solutions

Pitfall 1: Temporal Data Leakage

Problem: Epoch-level CV artificially inflates performance Solution: Always use subject-level (LOSO) cross-validation

Pitfall 2: Class Imbalance

Problem: Unequal epoch counts per subject Solution: Use subject-level majority voting, not epoch-level averaging

Pitfall 3: Overfitting to Noise

Problem: High-dimensional scattering coefficients Solution: Apply FDR correction, use regularization (e.g., L1 in SVM)

Pitfall 4: Ignoring Temporal Structure

Problem: Treating epochs as independent Solution: Use subject-level predictions, report subject-level metrics

Related Tools

• Kymatio: Python library for Wavelet Scattering Transform
• MNE-Python: EEG/MEG preprocessing and analysis
• SHAP: Model-agnostic explainability
• scikit-learn: Classification and cross-validation

References

• Original paper: arXiv:2607.05282
• WST theory: Mallat, 2012
• Schizophrenia EEG review: Herrmann & Demiralp, 2005
• Cross-frequency coupling: Canolty & Knight, 2010

Activation Triggers

Use this skill when:

• Analyzing resting-state EEG for psychiatric biomarkers
• Applying wavelet-based feature extraction to neural signals
• Designing subject-independent classification pipelines
• Keywords: "wavelet scattering", "schizophrenia", "EEG biomarker", "cross-frequency coupling", "LOSO cross-validation"