EEG-based Emotion Recognition in Resource-Constrained Environments: Reproducibility and Lightweighting of a Gompertz Fuzzy Ensemble Model on TorchEEG

EEG-based emotion recognition for resource-constrained devices is often challenged by irreproducible evaluation pipelines and oversized deep models. This study establishes a TorchEEG-based reproducible workflow on DEAP, employing trial-wise grouped 5-fold cross-validation to prevent trial-level information leakage and ensure consistent benchmarking. Building on this pipeline, we propose a lightweight multi-branch ensemble featuring a shared depthwise-separable 1D-CNN backbone with three shallow complementary branches (LSTM, GRU, and convolutional pooling), enabling parameter reuse with minimal overhead. For decision fusion, a Gompertz-function-based sample-wise adaptive fuzzy weighting is introduced to nonlinearly map branch confidence into normalized weights, improving robustness without additional heavy computation. On valence binary classification, the proposed model achieves 82.28% ± 0.58% accuracy with loss 0.588 ± 0.023 using only \((4.9\times10^4)\) parameters, delivering an approximately 33× parameter reduction versus commonly used million-parameter CNN baselines while maintaining comparable accuracy. Fusion ablations further indicate that the proposed weighting mitigates the large and unstable test loss observed with max-confidence selection. All experiments are conducted under CPU-only inference and evaluation settings to reflect practical deployment constraints.