A hybrid neural deep belief networks-gated recurrent unit algorithm with a leakage-aware evaluation protocol for ECG-HRV-based cognitive load classification

Cognitive underload and overload can impair knowledge workers’ performance. Electrocardiogram-derived heart rate variability (HRV) offers a practical signal for assessing cognitive demand. However, sample-wise evaluation may reward participant recognition rather than cognitive-state classification. This study proposes and evaluates a hybrid deep belief network–gated recurrent unit (DBN–GRU) algorithm on WESAD and SWELL. The model combines restricted Boltzmann machine pre-training with a gated classifier head. The primary WESAD evaluations use participant-disjoint protocols and validation-based model selection. Under leave-one-subject-out validation, the model achieves 68.4% accuracy, 62.8% balanced accuracy and 0.614 macro-F1. A majority-class predictor achieves 52.8% accuracy and 0.230 macro-F1. Participant-relative features improve performance for 13 of 15 participants ( p = 0.0015), whereas adversarial suppression of participant identity does not improve classification. Participant-relative features require unlabelled calibration data from each new participant. None of the eight alternative architectures differs significantly from the proposed model. Zero-shot cross-dataset transfer falls below the target majority-class baseline in both directions. Fine-tuning underperforms direct training on the target dataset. These results indicate a larger effect of participant-relative features than of architectural choice under the evaluated conditions as a quality assurance metrics.