Subject-Disjoint, Imbalance-Aware Gameplay–Rest Classification in Pediatric Mobile EEG Using Enriched Topological Features
Discuss this preprint
Start a discussion What are Sciety discussions?Listed in
This article is not in any list yet, why not save it to one of your lists.Abstract
Objective
Classifying cognitive state from pediatric mobile electroencephalography (EEG) recorded during naturalistic behavior is challenging under severe class imbalance, subject variability, motion artifacts, and low channel count. To avoid inflated performance estimates, this study enforces subject-disjoint evaluation and imbalance-aware metrics to distinguish Minecraft gameplay from eyes-open rest.
Methods
We propose Enriched Topological Features (ETF), which combine persistence landscapes with temporal entropy aggregation within Takens phase-space embeddings derived from persistent homology. ETF was evaluated on four-channel Muse EEG from 43 children using subject-disjoint five-fold GroupKFold cross-validation. Class imbalance was handled exclusively within training folds using synthetic minority oversampling or within-subject random undersampling, while test folds preserved natural class proportions. Sensitivity of topological parameterization was assessed across persistence landscape resolution and truncation depth using linear mixed-effects modeling.
Results
Under oversampling-based training, ETF achieved test balanced accuracy up to 65.39% and macro-F1 up to 64.48%, with minority-class (rest) recall reaching 49.51% using only TP9/TP10 electrodes. Relative to persistence-landscape-only representations, ETF improved macro-F1 by approximately 6–10 percentage points under subject-disjoint, imbalance-aware evaluation. Subject-level bootstrap analysis confirmed that performance trends were not driven by unequal epoch counts per subject. End-to-end processing required 676 ms per 4.5 s EEG epoch on CPU, supporting near–real-time feasibility.
Conclusion
ETF enables imbalance-aware gameplay–rest discrimination from low-density pediatric mobile EEG under subject-disjoint evaluation.
Significance
These results demonstrate that enriched topological representations remain informative under realistic motion and hardware constraints, supporting coarse engagement–rest state monitoring in pediatric mobile neuroengineering settings.
