Configurable Modular EEG Classification Framework with Multiscale Features and Ensemble Learning: A Reproducible...

Area of Science:

• Neuroscience
• Machine Learning
• Computational Psychiatry

Background:

• Deep learning models for EEG-based mental disorder classification are computationally intensive and lack interpretability.
• This limits their reproducibility and clinical deployment, especially in resource-limited settings.

Purpose of the Study:

• To propose a configurable, modular, and interpretable machine learning framework for EEG-based classification.
• To establish a benchmark for reproducible EEG analysis using schizophrenia detection as a use case.
• To rigorously evaluate the impact of different validation strategies on model generalization.

Main Methods:

• Developed a framework integrating standardized preprocessing, multiscale feature extraction, and minimum redundancy-maximum relevance feature selection.
• Implemented configurable ensemble learning and supported multiple validation strategies (random splits, k-fold cross-validation, leave-one-subject-out).
• Evaluated on two open EEG datasets (Warsaw IPN and Moscow adolescent cohort).

Main Results:

• Validation strategy significantly impacts model performance, with k-fold cross-validation overestimating accuracy.
• Leave-one-subject-out (LOSO) validation yielded substantially lower, more realistic performance metrics.
• Epoch-level accuracies ranged from 70.71% to 98.06%, and subject-level accuracies from 77.38% to 82.14% depending on the dataset and validation method.

Conclusions:

• Subject-independent evaluation is crucial to avoid performance overestimation due to data leakage.
• The proposed framework offers a low-complexity, interpretable, and extensible benchmark for reproducible EEG machine learning.
• The framework's interpretable features and modular design support broader neuroengineering and clinical decision-support applications.