Can Your Wearable Tell the Difference Between Excitement and Anxiety?

This piece covers what electrodermal activity (EDA) and related arousal sensors on wearables can and cannot distinguish about emotional states, based on published EDA and psychophysiology research. It does not cover heart rate variability algorithms, proprietary stress scores, or clinical diagnosis.

The foundational model of emotion that wearable stress algorithms are built on holds that arousal and valence are separate dimensions, and that skin conductance and heart rate track arousal only, not whether the feeling is good or bad. That means a wearable spike before a rollercoaster and a spike before a panic attack can look identical on the raw signal, because the sensor was never designed to read the difference.

The watch buzzes. Is that a good sign or a bad one?

There's a specific moment a lot of wearable users hit: the device flags an elevated stress reading right before something exciting, like walking into an interview that actually feels promising, or waiting for a flight that means a vacation, not a funeral. The number on the screen doesn't seem to care that the moment felt good. That mismatch is where the real question sits, not 'is my stress score accurate' in some abstract sense, but whether the underlying sensor has any way of knowing the emotional flavor of what it's picking up.

The answer sits in how electrodermal activity was defined as a research tool in the first place, and it's a more settled answer than the shifting number on a wrist display might suggest.

So what is the wearable actually measuring?

Electrodermal activity works by picking up tiny changes in skin conductance driven by sweat gland activity, which is controlled by the sympathetic nervous system, the branch of the nervous system that ramps up during any kind of arousal, not just the unpleasant kind. A wearable device putting a number on that signal is reporting the size and pattern of that sympathetic response. Research using this same signal to detect calm versus distress states has shown it can separate a relaxed state from an aroused one with reasonable accuracy in controlled testing using standardized emotional images. Separate research has used wrist-worn EDA to flag rising stress before surgery, a context that is unambiguously negative in valence.

Neither of those studies needed to ask whether the arousal was pleasant, they were built to detect the presence and intensity of a sympathetic response. That's a different question than the one a reader is really asking when they look at a spike before a first date and wonder if it means the same thing as a spike before a difficult meeting.

I think of this as closely related to a separate but distinct question, whether the stress number itself is measuring stress in any meaningful sense at all, which is covered in more depth in a companion piece on what wearable stress scores are actually built from.

Does the research offer any way around the arousal-only limit?

Some newer work is trying to sharpen what EDA alone can distinguish. One study built a framework to correct EDA readings for skin temperature effects in virtual reality settings, since heat can distort conductance readings independent of any emotional response, and found that temperature-adjusted features improved the distinction between stress-related and thermally-driven skin conductance patterns. That's a real methodological advance, but it's aimed at separating true stress signal from a thermal artifact, not at separating positive arousal from negative arousal within genuine emotional responses.

Other research has pushed on reliability rather than valence. A review applying generalizability theory to psychophysiological modalities, including EDA, examined how consistent these measures are across trials, tasks, and sessions, which matters for whether a single reading can be trusted at all before asking what emotional category it belongs to. Separate research has focused specifically on cleaning motion artifacts out of ambulatory EDA data, since a device worn on the wrist during daily movement picks up noise that has nothing to do with any emotional state, positive or negative.

None of this body of work resolves the excitement-versus-anxiety question, because none of it was designed to. The improvements are about signal cleanliness and measurement reliability. Adding a valence dimension to a sensor defined, from Russell's original framework onward, as an arousal-only channel was never the goal.

What EDA has been shown to flag, and what that says about the excitement question

EDA's track record outside the lab leans heavily toward flagging negative or clinically relevant states. A systematic review of electrodermal activity research found the studies were fairly consistent that a blunted, hypoactive electrodermal response is an established feature in depression, and found preliminary evidence that EDA monitoring may help differentiate phases of mood disorders and, in some studies, distinguish acutely suicidal patients from depressed patients who are not severely suicidal. Separate research examining diurnal EDA patterns from free-living wearable data in a large community sample found differences in electrodermal, temperature, and heart rate patterns associated with mental health status.

Both of those bodies of work are about detecting the presence of dysregulated or elevated arousal tied to negative mental health states. Sorting a given arousal spike into 'good nerves' or 'bad nerves' in the moment is a different task entirely, and not one either was built for. Guidelines for publishing EDA research have long emphasized that the signal reflects sympathetic arousal integrated with emotional and cognitive states broadly, not a labeled emotional category, and that framing hasn't changed as the sensors have moved from lab electrodes to wrist-worn consumer devices.

There's also evidence that different parts of the body can carry somewhat separate stress signals. Research combining brain imaging with machine learning found that skin conductance responses and heart rate responses to social threat were predicted by partly overlapping but largely distinct brain activity patterns, suggesting these two commonly tracked wearable metrics aren't just two windows onto the identical internal state. I find that distinction may matter for future sensor design, but it doesn't currently give a wearable a way to read valence either.

Common questions