Raw ECG representations achieved the highest macro-F1 and balanced accuracy (0.865) for classifying low versus high cognitive workload compared to conventional HRV-RF models (0.648).
Observational (n=17)
Can ECG-only representations accurately classify cognitive workload states during laparoscopic training?
ECG-only representations, particularly raw ECG, can effectively classify cognitive workload states during controlled laparoscopic training.
Absolute Event Rate: 0.865% vs 0.648%
Electrocardiography (ECG)-only workload-state classification offers a lower-burden physiological sensing route than denser multimodal, multi-sensor physiological, or neuroimaging setups for controlled laparoscopic training research. This study evaluated whether ECG-only representations can classify condition-derived cognitive workload states during a controlled laparoscopic peg transfer task performed under Control and auditory N-back conditions (N0, N1, and N2). Twenty surgical trainees from an advanced surgical-skills course completed the task protocol, and 17 participants entered ECG modelling after ECG quality-control exclusions. The retained ECG modelling dataset comprised 268 task blocks (Control/N0/N1/N2: 68/68/68/64), evaluated as held-out task-block predictions in a known-participant four-fold leave-one-round-out (LOTO) evaluation. Branch-specific raw ECG windows, recurrence-plot sequences, and heart-rate/time-domain heart-rate-variability inputs are detailed in the Methods, and model metrics were computed after reduction to task-block predictions. The primary endpoint was Surgery Task Load Index (SURG-TLX)-aligned low/high workload, defined as Control plus N0 versus N1 plus N2. Four-class and three-level endpoints were retained as secondary views. Raw ECG, recurrence-plot (RP)-derived, hybrid score-level fusion, and conventional heart-rate/time-domain heart-rate-variability Random Forest (HRV-RF) models were compared using a locked evaluation protocol, leakage-aware train-fold-only preprocessing, participant-clustered confidence intervals, and planned paired tests for the primary endpoint. On the primary low/high endpoint, raw ECG achieved the highest macro-F1/balanced accuracy (0.865/0.865), followed by the hybrid branch (0.847/0.847) and HRV-RF (0.648/0.649). Raw ECG and hybrid were supported over RP-derived and HRV-RF under the planned paired tests. On selected secondary endpoints, hybrid achieved higher macro-F1 values than raw ECG, consistent with possible endpoint-dependent RP-derived complementarity rather than a general hybrid advantage. These findings support ECG-only block-level workload-state classification in this controlled training setting. The evidence is retrospective and based on held-out rounds from known participants rather than subject-independent or deployment validation.
Jin et al. (Sun,) conducted a observational in Cognitive workload during laparoscopic training (n=17). Raw ECG representations vs. Conventional heart-rate/time-domain heart-rate-variability Random Forest (HRV-RF) models was evaluated on Surgery Task Load Index (SURG-TLX)-aligned low/high workload (Control plus N0 versus N1 plus N2). Raw ECG representations achieved the highest macro-F1 and balanced accuracy (0.865) for classifying low versus high cognitive workload compared to conventional HRV-RF models (0.648).