Why the study?
Efficient ECG emotion recognition requires addressing participant variability and the high computational cost of large random-kernel banks.
Population
12 WESAD participants and 20 DREAMER participants, with separate development cohorts
Comparison
250-group vs matched 5000-group DVSG-SFD configurations, and calibrated fusion vs CFF-DP
Design
Participant-held-out evaluation with separate development cohorts
Answers are grounded in this paper and cite their sources.
May enable efficient on-device ECG emotion analysis; leaves open clinical validation in cardiovascular cohorts.
Efficient electrocardiogram (ECG) emotion recognition must address participant variability and the cost of large random-kernel banks. We investigate dual-view stable group sequential feature detachment (DVSG-SFD), built on the Random Convolutional Kernel Transform (ROCKET). The method combines complete kernel-group pruning, source-participant rank dispersion and global window descriptors, with optional supervised calibration. Participant-held-out evaluation covers 12 WESAD participants over five seeds and 20 DREAMER participants over two seeds, with separate development cohorts. On WESAD, the 250-group source configuration achieves Macro-F1 0.4990 vs. 0.5061 for the matched 5000-group configuration, reducing the maximum kernel budget by 95%. With 40 labeled seconds per state for within-record calibration, fusion achieves a mean Macro-F1 of 0.7188 vs. 0.7219 for independently implemented CFF-DP. Under CPU-only execution, fusion achieves speedups of 2.07–2.22 for window processing and 53.2–63.8 for calibration computation across three WESAD participants, relative to traced CFF-DP inference and native fine-tuning, respectively. Calibration timing excludes label acquisition; inference process memory is higher. DREAMER source Macro-F1 scores remain below the uniform-random references. These results quantify the WESAD recognition–cost trade-off of compact random-kernel representations and supervised adaptation.
No takes yet. Share an insight, caveat, or question.
Liu et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: