PULSE: Privileged Knowledge Transfer from Electrodermal Activity to Low-Cost Sensors for Stress Monitoring

TL;DR

PULSE addresses the cost and unreliability of electrodermal activity (EDA) by using EDA solely during self-supervised pretraining to inject sympathetic-arousal information into cheap wearable sensors. The method learns modality-invariant shared embeddings and modality-specific private embeddings, aligning the shared space across sensors via a hinge loss and reconstructing inputs to prevent collapse. A frozen EDA encoder acts as a teacher to distill privileged EDA representations into deployable sensors (ECG, BVP, ACC, TEMP), enabling accurate stress detection at inference without EDA. Evaluated on the WESAD dataset with leave-one-subject-out evaluation, PULSE outperforms no-EDA baselines and symmetric alignment, and in some cases even surpasses the full-sensor baseline due to regularization and stable teacher guidance. The work demonstrates a practical path to cheaper, robust wearable stress monitoring by compressing privileged EDA information into accessible sensors, with broad potential for cross-domain transfer and future refinements.

Abstract

Electrodermal activity (EDA), the primary signal for stress detection, requires costly hardware often unavailable in real-world wearables. In this paper, we propose PULSE, a framework that utilizes EDA exclusively during self-supervised pretraining, while enabling inference without EDA but with more readily available modalities such as ECG, BVP, ACC, and TEMP. Our approach separates encoder outputs into shared and private embeddings. We align shared embeddings across modalities and fuse them into a modality-invariant representation. The private embeddings carry modality-specific information to support the reconstruction objective. Pretraining is followed by knowledge transfer where a frozen EDA teacher transfers sympathetic-arousal representations into student encoders. On WESAD, our method achieves strong stress-detection performance, showing that representations of privileged EDA can be transferred to low-cost sensors to improve accuracy while reducing hardware cost.

Figures (7)

• Figure 1: The PULSE Framework. Our framework uses privileged knowledge transfer from a frozen EDA encoder to students built on low-cost sensors. In the pretraining stage, student encoders learn modality-invariant shared embeddings alongside modality-specific private embeddings. Knowledge transfer is then achieved by aligning the students' shared embeddings with the privileged EDA teacher. Finally, during finetuning, the learned embeddings are used for supervised stress detection without requiring EDA at inference.
• Figure 2: Pretraining Setup. In pretraining, each of the cheap sensor encoders outputs shared embeddings (colored boxes) and private embeddings (white boxes). The shared embeddings are aligned across modalities via a hinge loss objective, then averaged into a single shared embedding (magenta boxes). This averaged shared embedding, together with the private embeddings, is fed into the decoder for signal reconstruction. The EDA MAE is trained separately via only reconstruction loss.
• Figure 3: Learning curve showing anchored contrastive loss in early experiments
• Figure 4: Effect of adding reconstruction during KD. Without reconstruction, shared embeddings collapse (cosine $\approx 1.0$, near-zero variance). Adding reconstruction restores variance by several orders of magnitude, preventing collapse and enabling meaningful shared geometry.
• Figure 5: Convergence behavior in pretraining