The Science Behind Wearable Devices That Monitor Anxiety and Depression

A wristband measures the interval between two heartbeats.

A diagnosis of generalized anxiety disorder requires six months of excessive worry the person struggles to control, plus at least three of six specified symptoms, plus clinically significant distress or impairment.

Nothing in that first sentence gets you to the second one on its own, and that gap is the central scientific problem in wearable mental health monitoring.

You will discover these devices in your reading, in your placements, and eventually in your patient's pockets. Some of the research behind them is genuinely strong. Some of it is a correlation wearing a screening tool's clothes. Telling the two apart is a skill you can start building now, and it begins with knowing what the hardware on someone's wrist is physically able to notice.

What a Wristband Can Actually Detect

Consumer and research-grade wearables sense a short list of physical quantities. Everything else is inference layered on top:

• Heart rate and heart rate variability, usually from photoplethysmography: a green LED shines into the wrist and a photodiode reads how much light the pulsing blood absorbs.
• Electrodermal activity, the sweat-gland conductance shift that appears a second or two after a startle.
• Movement, from a three-axis accelerometer, which is also how the device infers sleep timing and restlessness.
• Skin temperature, and on some devices a respiration estimate derived from chest or wrist motion.

That is the raw material. The path from those signals to anxiety runs mostly through the autonomic nervous system: sympathetic arousal speeds the heart and flattens its variability, while parasympathetic tone does the reverse. The evidence for that link is better than most students expect. Pooling 36 studies covering 2,086 patients with an anxiety disorder and 2,294 controls, Chalmers and colleagues found reduced high-frequency heart rate variability at an effect size of Hedges' g = -0.29, alongside a larger reduction of -0.45 in time-domain measures.

Read that effect size slowly, because the whole field pivots on it. Small-to-moderate at the group level means the two distributions overlap heavily. Plenty of people with panic disorder sit comfortably inside the healthy heart-rate-variability (HRV) range, and plenty of unbothered people sit below it. The same analysis found no HRV reduction at all in obsessive-compulsive disorder, which should tell you something useful: "anxiety" is not one physiological thing, so no single autonomic number can stand in for all of it.

From Raw Signal to a Number a Clinician Can Trust

Between the photodiode and anything a clinician would act on sits a processing pipeline, and most of the failure modes live there rather than in the sensor:

1. Sampling and artifact rejection. Wrist photoplethysmography is degraded by motion, cold hands, tattoos, and loose straps. Some fraction of every dataset is unusable, and the rule a team picks for discarding bad windows quietly shapes the result.
2. Feature extraction. Raw beat intervals become RMSSD, SDNN, frequency-band power, sleep-fragmentation indices. Every choice here embeds an assumption about what matters.
3. Modeling. Those features get mapped onto an outcome, which in practice is almost always a questionnaire score rather than a clinical interview.
4. The claim. Someone decides what the output is permitted to say: "your readiness is low" or "consider screening for depression."

Step four is where regulators show up. A vague wellness score sits outside their remit. The moment a device output is intended to inform a diagnosis or a treatment decision, the software itself becomes a regulated medical device in both the US and the EU, and the engineering discipline changes shape.

You get version-controlled requirements, traceable risk analysis, validation documented against a clinical reference standard, and a defined process for what happens when the model gets retrained. That shift from ordinary analytics to medical device software development explains why so few consumer wearables make explicit claims about mental illness, even when their own research teams have published papers suggesting they could.

The regulatory route does get used. In April 2024 the FDA cleared Rejoyn, developed by Otsuka and Click Therapeutics, as the first prescription digital therapeutic authorized for the adjunctive treatment of major depressive disorder symptoms in adults 22 and older already taking antidepressants.

It is a six-week app-based cognitive-emotional training program rather than a sensor, but it set a visible benchmark for the evidence a mental health software claim has to carry.

Why Lab Accuracy Doesn't Survive Contact With Real Life

Published performance figures look excellent. A meta-analysis in npj Digital Medicine reported pooled peak accuracy of 0.89, sensitivity of 0.87, and specificity of 0.93 for wearable-based AI detecting depression. Then read the discussion, where the authors state plainly that heterogeneity between studies, a predominance of retrospective designs, and heavy reuse of the same public datasets all limit how far those numbers travel.

Three specific things go wrong on the way out of the lab.

Base rates dilute good specificity

Work the arithmetic yourself. Take that 0.93 specificity and 0.87 sensitivity, and assume for the sake of the example a screening population where 10% of people currently meet criteria for major depression. In 1,000 people, 100 have it and 87 get flagged correctly. Of the 900 who do not, 63 get flagged anyway. Positive predictive value lands near 58%, close to a coin flip despite an impressive-looking confusion matrix. Lower the assumed prevalence and it gets worse fast, which is exactly why population-wide passive screening is a harder sell than screening people who already presented with a complaint.

Subject leakage inflates everything

Many models are validated on held-out days drawn from the same people the model trained on. That design measures whether the algorithm has learned one participant's personal baseline, not whether it has learned depression. Splitting by subject is the only version that answers the clinical question, and it usually costs several points of accuracy.

Almost nobody keeps wearing it

A panel-based analysis of real-world mental health app engagement found a median 15-day retention of just 3.9%. A model that needs fourteen consecutive days of clean sensor data to flag a coming depressive episode has an obvious problem if the typical user has already stopped by day ten.

How to Read One of These Studies Without Getting Fooled

Next time a wearable mental health paper lands on your reading list, work through this before you accept the headline number:

• What was the ground truth label? A PHQ-9 self-report is not a structured clinical interview, and a model that predicts questionnaire scores has predicted questionnaire scores.
• Was the train/test split by subject or by day? If the paper does not say, assume the weaker option.
• Prospective or retrospective? Retrospective analyses of an existing dataset invite quiet re-specification until the numbers cooperate.
• Is positive predictive value reported alongside sensitivity and specificity? If not, compute it yourself using the sample's actual base rate.
• Who dropped out, and were they different from those who stayed? Attrition in mental health studies is rarely random, and the people who stop wearing the device are often the ones doing worst.
• What is the comparison condition? Beating chance is easy. Beating a two-item PHQ-2 screener that takes twenty seconds and costs nothing is the bar that matters clinically.

None of this means the technology is empty. Passive sensing genuinely can catch things self-report misses, particularly sleep disruption and social withdrawal, and it captures them continuously instead of at a single appointment. The honest current position is that wearables are a promising source of supplementary signal, not a diagnostic instrument.

Frequently Asked Questions

Can a smartwatch diagnose anxiety or depression?

No, and no device on the market is cleared to. The closest legitimate claims are risk flags or symptom-tracking support, always alongside clinician assessment.

Is HRV a real biomarker for anxiety?

It is a real group-level association with a small-to-moderate effect size and meaningful overlap between clinical and healthy ranges. That makes it useful for research and unreliable for judging any individual.

Should I cite consumer wearable data in my own research?

You can, provided you validate the device against a reference measure for your specific outcome and report which firmware and algorithm version produced the data. Vendors change their algorithms without notice, which can break comparability between cohorts collected months apart.

What to Take From This

Wearables measure autonomic and behavioral proxies, not psychological states, and the distance between those two things is where the science lives. Three things are worth keeping: group-level effect sizes as modest as g = -0.29 cannot support individual diagnosis; reported accuracy means little until you know the base rate and the validation split; and a regulated clinical claim demands a level of documented validation that a wellness score never has to meet.

When you next read one of these papers, start with the ground truth label. Most of what you need to judge the work is right there.