This is the authors' abstract. We don't add key points for this paper.
Electrodermal activity (EDA) is a widely used marker of sympathetic arousal and stress, in both daily-life and clinical monitoring, but EDA signals from wearable devices are frequently degraded by artifacts, and expert-labeled data for quality assessment are scarce. We present a lightweight EDA signal quality index (SQI) system that uses unsupervised pre-training to reduce reliance on labeled data. The model is first pre-trained on unsupervised denoising and forecasting tasks, then fine-tuned using a publicly available expert-labeled dataset. On the EDABE benchmark, our method achieves a mean ROC-AUC of 0.851 and a mean AUPRC of 0.494, improvements of 8% and 18.5%, respectively, over the strongest baseline. The model also generalizes across devices, body locations, and types of motion: applied zero-shot to AffectiveROAD, wrist-worn Empatica E4 recordings from real-world driving (whereas training used finger-worn sensors), it achieves a ROC-AUC of 0.740, a 7.9% improvement over the same network trained from scratch, using high wrist acceleration as the reference for motion artifacts. The model requires 3.9 MFLOPs per inference. These results show that unsupervised pre-training improves performance, enabling scalable, low-power EDA quality assessment for continuous wearable monitoring.
Answers are grounded in this paper and cite their sources.
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: