ABSTRACT Dependable transfer of brain signals from motor imagery EEG must adhere to strict latency and memory constraints while maintaining accuracy in the face of noise and drift. A hybrid Echo State Network‐Long Short‐Term Memory (ESN‐LSTM) pipeline is shown here. This pipeline combines robust preprocessing with automatic time‐lag alignment between predictions and targets. To capture structure lost by linear errors alone, the evaluation combines traditional regression metrics (MSE/MAE/ R 2 ) with nonlinear dependence measures (time‐resolved distance correlation and HHG omnibus testing). A leave‐one‐subject‐out (LOSO) procedure is used to investigate cross‐subject generalization, and multi‐signal‐to‐noise ratio (SNR) stress tests are conducted to evaluate robustness. In ablation experiments, the impact of filtering, normalization, alignment, and important hyperparameters (reservoir size/spectral radius/leak; LSTM layers/hidden/dropout) is isolated. On the other hand, an efficiency snapshot reports latency and RAM usage under identical workloads for ESN‐Only, LSTM‐Only, and ESN‐LSTM modes. Across all participants, the hybrid consistently improves explained variance and dependence scores while maintaining a controlled computational cost, indicating that it is feasible for near‐real‐time use. All tables and Figures are regenerated from logged CSVs using scripts, configuration files, and fixed seeds. This ensures that reproducibility is maintained.
Alzubaidi et al. (2026) studied this question.