Can Wearables Flag Mood Changes Before You Feel Them?

A wearable may notice that your sleep has shifted before you recognize a mood shift yourself. That is the most credible promise in passive mental-health sensing. It is not a diagnosis, and it is not proof that an app can forecast your next episode. The strongest evidence points to ordinary-looking measures: when you fall asleep, how much that timing varies, how active you are during the day, and how stable your 24-hour routine remains.

The distinction matters because flashy claims travel fast. A 2025 deep-learning report claimed that a 1- to 1.5-hour heart-rate-variability (HRV) recording could distinguish schizophrenia from bipolar disorder with 80% to 83% accuracy. That sounds clinical. It is not ready for clinical use. A small sample, overlapping windows from the same people, no independent external validation, and uncontrolled psychiatric medication can all make a classifier look smarter than it is. The broader evidence supports a more modest conclusion: passive sensors can track behavioral changes associated with psychiatric symptoms, especially changes in sleep and daily activity.

The useful signal is circadian regularity

A 2025 systematic review of 45 bipolar-disorder digital-biomarker studies found the most repeated signals in combined sleep and activity data. Depressive states tended to coincide with reduced mobility, later and more variable sleep timing, and lower daytime light exposure. Hypomania and mania tended to coincide with shorter, earlier sleep and higher, more erratic activity. These are associations with mood state, not tests that establish its cause.

This pattern makes clinical sense without turning a wristband into a psychiatrist. Sleep disruption and altered activity already form part of the observable picture of mood episodes. Actigraphy gives those changes a daily record instead of relying only on memory at a clinic visit. The review reported machine-learning AUROCs of 0.80 to 0.98 in larger studies, but those figures do not establish real-world performance. All 15 machine-learning studies in the review lacked independent external validation. A model can perform well on data that resemble its training set and fail when it meets a different clinic, device, medication mix, or population.

The same review found that 64% of studies enrolled 50 or fewer participants, 49% monitored people for 100 days or less, and 29% collected data only in clinical settings. Bipolar depressive episodes can last months. A short study may miss onset, progression, and recovery—the sequence a relapse-prediction tool would need to learn.

Why daily data may help across diagnoses

The signal does not appear confined to one diagnostic label. The longitudinal DeeP-DD study used consumer-grade GENEActiv wristbands in people with schizophrenia, bipolar disorder, depression, and anxiety. It found that sleep and activity patterns could function as transdiagnostic behavioral biomarkers, with daily-resolution features more actionable than weekly or monthly summaries. The study supports feasibility in real-world psychiatric populations; it does not show that a wearable alert improves treatment or prevents an episode.

A 2022 systematic review of actigraphy, mood instability, and impulsivity also found bidirectional associations between circadian disruption and both mood instability and impulsivity across psychiatric diagnoses. "Bidirectional" is a warning against simple stories. A disrupted rhythm may accompany worsening mood, contribute to it, result from it, or all three. A wrist sensor measures movement and timing. It cannot settle that causal question.

Actigraphy changes also appear early in illness trajectories. In young people seeking mental-health care, a 2025 study of clinical stages and pathophysiological subgroups found systematic differences in actigraphy-derived circadian rhythms, sleep-wake patterns, and physical activity. That makes passive monitoring interesting for early-warning research. It does not create a validated threshold for identifying an individual's diagnosis or future course.

Smartphones add context, then run into personalization

Phones can add information that a wrist device cannot: location patterns, accelerometry, screen time, app use, calls, and typing behavior. A 2025 systematic review of longitudinal mobile sensing for depression found that passive data could distinguish depression-severity states within individuals during monitoring periods longer than three months. Sleep and physical activity were the primary clinically relevant features.

But the review identified the problem that accuracy headlines often bury: a predictor at the group level can point the opposite way for an individual. For example, a population-level association between sleep duration and worsening depression may reverse within a particular person. That makes a population-trained model a poor substitute for a personal baseline. An app trained on thousands of users may be wrong in a highly customized way.

A 2024 systematic review of smartphone digital phenotyping found that continuous phone data could detect behavioral patterns related to stress, anxiety, and mild depression. Feasibility is real; clinical utility remains unproven. The field still needs studies in which sensor-triggered alerts change care and measurably improve patient outcomes.

HRV is real physiology, not a psychiatric fingerprint

HRV measures variation in the time between heartbeats. It reflects autonomic regulation, and it changes with sleep, exercise, illness, medication, alcohol, and measurement conditions. A 2025 umbrella review of meta-analyses found lower HRV across several psychiatric conditions, but also high heterogeneity tied to recording context, recording duration, medication status, and psychiatric state at measurement.

That heterogeneity is fatal to claims that a short HRV recording can diagnose a specific psychiatric condition. Antipsychotics, mood stabilizers, and SSRIs can alter HRV independently of symptoms. Exercise can alter it, too: a systematic review and meta-analysis in people with type 2 diabetes found that exercise training improved HRV measures. A responsive biomarker can be useful for studying physiology, but it becomes hard to interpret when many inputs move it at once.

Windowed HRV analyses create another trap. If researchers cut one recording into many heavily overlapping segments, they can generate thousands of observations from a small number of people. Those segments are not independent data points. Unless the analysis keeps participants separate during training and testing, apparent precision can become a mathematical echo of the same recording.

Consumer wearables measure imperfectly

Consumer devices are convenient, but convenience does not equal a calibrated medical instrument. A 2025 systematic review of consumer wearables in severe mental illness found sleep and physical activity to be the most replicated predictors, while noting that studies rarely provided sensitivity and specificity thresholds that clinicians could act on. Device error also matters: a 2024 systematic review of the ActiGraph GT9X Link found that its step estimates generally underestimated steps and varied by activity type and individual characteristics.

Missing data are not necessarily random, either. A person may stop wearing a device during a depressive episode or remove it during agitation. The bipolar-disorder review noted that researchers rarely model this missingness, even though it may contain behavioral information. Ignoring it can bias a model; treating it as a diagnosis would be worse.

What to do with your own data

If you have a personal or family history of mood instability, the most evidence-based use of a wearable is simple: track your own sleep timing, night-to-night variability, daytime activity, and routine stability over weeks or months. Look for changes from your usual pattern, ideally alongside a brief mood and medication log. A single bad night is a human experience, not a diagnostic event.

• Use daily data to establish a personal baseline rather than chasing population averages.
• Pay attention to sustained shifts in sleep timing and daytime activity, especially if they coincide with mood, energy, or judgment changes.
• Discuss meaningful changes with a clinician who knows your history; do not change medication because a dashboard changed color.
• Treat HRV trends as nonspecific physiological context, particularly if medication, training, illness, alcohol, or sleep has changed.

There is also a metabolic reason to care about regularity. In a cohort of 225 young people recruited from early-intervention mental-health services, a 2025 observational study linked disrupted 24-hour rest-activity patterns with immune-metabolic risk factors in people with mood disorders. The finding is an association, not evidence that correcting a wearable-derived rhythm score prevents cardiovascular disease. Still, a stable sleep-and-activity routine is a sensible behavior to observe, because it is relevant to both mental health and everyday metabolic health.

Wearables can supply useful early-warning context. They cannot currently supply a psychiatric diagnosis, a validated relapse prediction, or proof that an alert will improve an outcome.

The practical conclusion is deliberately unglamorous: use passive data to notice persistent changes in your own routine, then bring those observations to real clinical judgment. The research has earned that level of confidence. It has not earned more.