⚡ electrical engineering

Unit-Independent Low-Rate Wrist GSR Processing for Stress Detection Using Phasic nSCR Features

This paper proposes a unit-independent, low-rate wrist GSR processing pipeline that extracts phasic nSCR/min features to achieve high-accuracy stress detection comparable to higher-frequency palmar measurements, demonstrating the feasibility of using wearable devices for reliable stress monitoring.

Original authors: Zequan Liang, Sally Hang, Geneva M. Jost, Ning Miao, Wei Shao, Mahdi Pirayesh Shirazi Nejad, Hossein Sayadi, Ehsan Kourkchi, Setareh Rafatirad, Camelia E. Hostinar, Houman Homayoun

Published 2026-07-10

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Zequan Liang, Sally Hang, Geneva M. Jost, Ning Miao, Wei Shao, Mahdi Pirayesh Shirazi Nejad, Hossein Sayadi, Ehsan Kourkchi, Setareh Rafatirad, Camelia E. Hostinar, Houman Homayoun

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine your skin is like a tiny, invisible radio station that broadcasts your stress levels. When you get nervous, your sweat glands turn on the volume, sending out little "pulses" of electricity. Scientists have long known this, but there's a catch: measuring these pulses from your palm (like a classic lab test) is easy because the signal is loud and clear. But trying to catch that same signal from your wrist? That's like trying to hear a whisper from across a noisy room. The wrist signal is often so quiet and different in size that standard tools get confused, thinking a whisper is a shout or missing it entirely.

This paper suggests a clever workaround: instead of trying to measure how loud the whisper is (which changes wildly between a palm and a wrist), let's just count how many times it whispers in a minute.

The Big Idea: Counting the Blips
The researchers, a team from UC Davis and California State University, built a special "translator" for wrist-worn devices. They collected data from 31 people wearing two gadgets at once: a high-tech wristband called "We-Be" and a lab-grade palm sensor called "MindWare." The participants did four things: sat quietly, stood quietly, talked about boring stuff (like making a sandwich), and then faced a stressful job interview simulation called the TSST.

The team realized that the raw numbers from the wrist and the palm were totally different—like comparing a whisper to a shout. So, they invented a new way to process the data. First, they cleaned up the noisy signal. Then, they used a mathematical trick called "robust z-score normalization." Think of this as a universal volume knob that turns every signal down to a standard scale, so a tiny wrist blip and a huge palm blip look the same size on the graph.

Once the signals were on the same playing field, they stopped looking at the height of the waves and started counting the peaks. They counted the number of skin conductance responses per minute (nSCR/min). It's like ignoring how loud a drummer is hitting the drum and just counting how many beats they play in a minute.

What They Found
The results suggest that this "counting beats" method works surprisingly well. When they used a computer program (called a Random Forest) to look at the wrist data, it could tell the difference between a calm person and a stressed person with a balanced accuracy of 0.823 (when comparing stress to sitting) and 0.871 (when comparing stress to standing). That's pretty good for a wristband!

Interestingly, the paper shows that you don't need super-fast data to get these results. The wristband recorded data at 100Hz (100 times a second), but the team tested what happened if they slowed it down to 25Hz (25 times a second). The results suggest that slowing it down didn't hurt the accuracy at all; in some cases, the slower 25Hz data performed just as well as the faster 100Hz data. This is a big deal because slower data means the battery lasts longer, which is great for wearable devices.

What This Doesn't Do
It's important to note what this study doesn't say. The paper explicitly argues against relying on the absolute size of the signal (amplitude) for wrist-based stress detection. They show that trying to use the same "loudness" rules for a wrist as you do for a palm leads to confusion. The wrist simply doesn't produce the same massive spikes as the palm, so trying to measure "how stressed" someone is based on signal height is a dead end.

Also, while the wristband did a good job, it wasn't perfect. The lab-grade palm sensor (MindWare) still did a better job at spotting stress than the wristband. The authors suggest that the wristband is promising but still has room to get better. They also found that the "neutral speaking" task (talking about sandwiches) was the hardest to tell apart from the stressful job interview, suggesting that even talking about boring things can trigger some stress responses that are hard to distinguish from real anxiety.

The Takeaway
In short, this paper suggests that we don't need a giant lab machine on our wrists to detect stress. By ignoring the confusing differences in signal size and just counting the number of stress "blips" per minute, we can get a pretty accurate picture of how stressed someone is. It's a step toward making stress-detecting watches that are cheap, battery-friendly, and actually work, even if they aren't quite as perfect as the expensive lab equipment just yet.

Technical Summary: Unit-Independent Low-Rate Wrist GSR Processing for Stress Detection

Problem Statement
Galvanic skin response (GSR), or electrodermal activity (EDA), is a standard physiological indicator for stress detection due to its reflection of sympathetic nervous system activity. However, deploying GSR monitoring on wrist-based wearable devices presents a significant challenge: the absolute amplitude of skin conductance measured at the wrist differs substantially from laboratory-grade measurements taken at the palm or fingers. This "wrist-palmar amplitude mismatch" arises from differences in sweat gland density and local skin properties across body regions. Consequently, traditional processing pipelines that rely on fixed amplitude thresholds (e.g., a minimum response of 0.05µS) for detecting skin conductance responses (SCRs) often fail when applied to wrist data, as thresholds appropriate for palmar measurements may be too strict or too sensitive for wrist signals.

Methodology
To address these challenges, the authors propose a unit-independent, low-rate wrist GSR processing pipeline designed to extract the number of skin conductance responses per minute (nSCR/min) as a stress-related feature. The methodology involves the following steps:

Data Collection: Paired GSR recordings were collected from 31 adult participants using two devices simultaneously: the We-Be wrist-worn device (placed on the nondominant wrist) and MindWare palm electrodes (placed on the nondominant palm). Participants completed four task segments: a 3-minute sitting baseline, a 3-minute standing baseline, a 5-minute neutral speaking task, and a 10-minute Trier Social Stress Test (TSST).
Signal Preprocessing:
- Cleaning: Raw GSR signals were filtered using a 2Hz low-pass filter and smoothed with a 1-second rolling median to reduce high-frequency noise.
- Decomposition: The cleaned signal was decomposed into tonic Skin Conductance Level (SCL) and phasic Skin Conductance Response (SCR) components using cvxEDA, a convex-optimization-based method implemented in NeuroKit2.
- Normalization: To mitigate the amplitude mismatch between wrist and palm, the phasic SCR signal underwent robust z-score normalization (Eq. 1), which uses the median and median absolute deviation rather than the mean and standard deviation.
- Peak Detection: SCR peaks were detected on the normalized signal using criteria independent of absolute units: a minimum peak prominence of 0.2, a minimum width of 1 second, and a minimum distance of 1 second between peaks.
- Feature Extraction: The count of detected peaks was normalized by the segment duration to calculate nSCR/min.
Classification: The nSCR/min feature was used as input for binary classification (TSST vs. non-stress conditions) using Logistic Regression, Random Forest, and a threshold-based classifier. Evaluation was performed using leave-one-subject-out (LOSO) cross-validation.
Sampling Rate Analysis: The study compared the performance of the original 100Hz We-Be data against downsampled 25Hz data to assess the viability of low-rate sensing.

Key Contributions

Dataset Creation: The authors constructed a dataset containing paired wrist and palmar GSR recordings from 31 participants across stress-inducing and baseline tasks.
Unit-Independent Pipeline: A novel processing pipeline was proposed that shifts the analysis focus from absolute conductance amplitude to phasic response rate (nSCR/min). By combining robust normalization with phasic peak detection, the pipeline effectively handles the wrist-palmar amplitude mismatch without requiring device-specific calibration.
Low-Rate Feasibility: The study demonstrated that downsampling wrist GSR data from 100Hz to 25Hz preserves sufficient information for segment-level stress detection, supporting the design of lower-power wearable systems.

Experimental Results

Feature Statistics: While raw GSR and SCL values showed substantial absolute range differences between the We-Be (wrist) and MindWare (palm) devices, the nSCR/min feature exhibited a consistent increasing trend from baseline tasks to the TSST across both devices. nSCR/min also achieved higher Pearson correlation coefficients between devices compared to raw GSR or SCL.
Statistical Separation: Paired comparisons showed that nSCR/min could statistically distinguish the TSST from sitting and standing baselines for both devices (Wilcoxon $p < 0.05$ ). The separation was strongest between TSST and baselines, and weakest between TSST and neutral speaking, suggesting the latter may also elicit stress-related responses.
Classification Performance:
- Using Random Forest on 25Hz We-Be data, the system achieved balanced accuracies of 0.823 (TSST vs. Sitting) and 0.871 (TSST vs. Standing).
- While the laboratory-grade MindWare system generally achieved higher performance (consistent with stronger palmar EDA), the wrist-based We-Be device still provided discriminative information for stress detection.
- Sampling Rate: The 25Hz downsampled features achieved comparable or even better balanced accuracy than the 100Hz features across all tasks and models. For instance, Random Forest accuracy for TSST vs. Standing improved from 0.823 (100Hz) to 0.871 (25Hz).

Significance and Claims
The paper claims that the proposed unit-independent, low-rate wrist GSR processing pipeline is feasible for wearable stress detection. By extracting nSCR/min, the method successfully captures stress-related phasic responses despite the significant amplitude differences between wrist and palm measurements. The results suggest that wrist-based wearables do not need to match the absolute amplitude accuracy of laboratory systems to detect stress; instead, they can rely on normalized response rates. Furthermore, the preservation of performance at 25Hz indicates that this approach can support low-power, continuous monitoring designs suitable for long-term wearable health applications. The authors modestly acknowledge that while the wrist device shows promise, there remains room to improve accuracy by addressing signal quality and motion noise in future work.

Technical Summary: Unit-Independent Low-Rate Wrist GSR Processing for Stress Detection

More like this