Studies show smartphone sensor data can flag depression and suicide risk with moderate accuracy. What the research found, its limits, and privacy risks.
Your smartphone records a steady stream of data about how you live: how far you move, when the screen turns on, how long your calls last, when the battery gets charged. Researchers call the practice of turning that data into health signals digital phenotyping, and a growing set of studies is testing whether it can flag changes in mental health before a person reports them.
The results so far are promising but modest. Models built on phone data reach moderate accuracy in research settings, and none of these tools is a diagnosis.
Digital phenotyping uses passive smartphone data (GPS, accelerometer, screen activity, battery level, call logs) and machine learning to estimate mental health risk. In a 2026 study of 103 London teenagers, models that combined passive data with daily self-reports reached balanced accuracies of 0.77 for suicidal ideation, 0.71 for high-risk SDQ scores and 0.70 for eating disorder risk.
The appeal is clear. Traditional mental health monitoring relies on self-report and scheduled appointments. Phone data is collected continuously and without extra effort from the patient.
The Science Behind Behavioral Fingerprints
Machine learning models trained on smartphone data look for patterns that track with mental health measures. A study first posted on arXiv in January 2025 and published in the Journal of Medical Internet Research in February 2026, recruited 103 adolescents with a mean age of 16.1 years from three UK secondary schools. Participants used the Mindcraft app for 14 days, giving daily self-reports and passive sensor data.
Using balanced accuracy (a measure that corrects for uneven group sizes), the combined models scored:
- 0.77 for suicidal ideation
- 0.71 for high-risk scores on the Strengths and Difficulties Questionnaire (SDQ)
- 0.70 for eating disorder risk
- 0.67 for insomnia
Models that combined active and passive data outperformed models using either source alone. The authors described the work as a feasibility study in a nonclinical group, so these figures are not ready for clinical use.
What Your Phone Actually Monitors
Passive data collection runs in the background. The studies below drew on several kinds of signals:
Movement and Location Patterns:
- GPS: How far and how widely a person moves
- Accelerometer data: Physical activity levels and movement patterns
- Location diversity: How many different places a person spends time
Communication Behavior:
- Call duration: Incoming and outgoing calls
- Text and call logs: Volume and timing of communication
Device Interaction:
- Screen events: How often the screen turns on and off
- App usage: Which apps are used and for how long
- Battery level: Charging patterns over the day
Temporal Patterns:
- Daily rhythms: When a person is active, on the phone or at rest
- Sleep estimates: Inferred in some studies from screen and movement data
In a 2024 study in the Journal of Medical Internet Research, researchers followed 164 participants, including patients with depressive disorders and healthy controls, for up to a year. The most important phone features for classifying depression were screen-off events, battery charge levels, communication patterns, app usage and location data. For a different approach to measuring depression, see our article on blood-based depression biomarkers.
The Adolescent Monitoring Study
A PLOS Digital Health study published in July 2025 is among the longest digital phenotyping efforts in young people. Researchers at Harvard and NYU followed 48 participants aged 14 to 19 (average age 15.85 years, 54% female, 54% from minoritized racial or ethnic backgrounds) for 18 months using the Beiwe app. The group included adolescents with bipolar disorder and peers with no mental health diagnosis.
Completion rates showed a clear pattern:
- 99% of clinical interviews completed (826 of 835 scheduled)
- 89% of days with passive data collected (22,233 of 25,029)
- 47% of thrice-weekly surveys completed (4,945 of 10,448)
Passive data collection stayed steady over time, while survey completion fell from 65% in the first 20 weeks to 30% in the final 19 weeks. Passive monitoring asks almost nothing of the participant, which is its main advantage.
"Results of this study suggest digital phenotyping has significant potential as a method of long-term mental health monitoring in adolescents. In contrast to traditional methods, including interview and self-report, it is lower burden and provides more complete data over time."
Huang, Emedom-Nnamdi, Onnela and Van Meter, PLOS Digital Health, 2025
Each participant generated an average of 4,101 MB of accelerometer data and 120 MB of GPS data over the study. The authors note that the next step is to test whether this data actually tracks changes in mental health.
The Depression Detection Paradox
The Mobile Monitoring of Mood (MoMo-Mood) study, published in JMIR Mental Health in February 2025, recruited 188 people: patients in a major depressive episode (with major depressive disorder, bipolar disorder, or borderline personality disorder) and healthy controls. Passive smartphone data and depression questionnaires were collected for up to 12 months, though most participants did not stay the full year.
Some findings were counterintuitive. Longer incoming calls and more variation in physical activity were linked to less severe depression. Longer outgoing calls were linked to more severe depression.
Healthy controls showed more diverse daily communication patterns than patients, and patients moved around less widely. In contrast, patients showed more varied patterns of smartphone use than controls. For related research on how early experiences shape the brain, see our article on childhood trauma and adult neural pathways.
In the 2024 JMIR study described above, a model using smartphone behavioral markers classified current depression with 82% accuracy and detected changes in depression status with 75% accuracy.
From Prediction to Prevention
The longer-term goal is early intervention. A 2023 study in the Journal of Medical Internet Research, part of the Spanish SmartCrisis project, followed 225 patients with a history of suicidal thoughts or behavior for six months. An algorithm built daily activity profiles from passive phone data (distance walked, steps, time at home and app usage) and flagged shifts in those profiles. Those shifts predicted suicide risk within a one-week window with an area under the curve of 0.78, which the authors describe as good accuracy.
Researchers see this as a route to "just-in-time" support. In principle, when a model flags elevated risk, a system could:
- Send a check-in prompt through a mental health app
- Alert a clinical team or a designated support contact
- Suggest coping strategies, such as those discussed in our article on gratitude and mental health
- Recommend booking an appointment with a provider
These uses are still being studied. In practice, clinicians would get a record of behavior between visits, rather than relying only on what a patient remembers at an appointment.
Privacy and the Passive Monitoring Dilemma
Digital phenotyping raises hard ethical questions. Location data can reveal where a person spends time and who they visit. Communication logs can reveal social withdrawal and relationship patterns.
In the PLOS Digital Health adolescent study, 94% of participants used iPhones. The authors noted that phone operating systems can affect data collection, for example because iOS may deactivate apps that are not used often. They also wrote that digital phenotyping involves sensitive data, including location, and that guidelines for this kind of data collection are not yet standardized.
As these tools move from research to clinical care, open questions include:
- Patient control over data sharing and deletion
- Transparency in how algorithms reach their predictions
- Protection against discriminatory uses
- Responsibility when a prediction is wrong or a warning sign is missed
Your phone already collects the raw material for this kind of monitoring. The research shows it can pick up some signals linked to depression and suicide risk, with moderate accuracy and mostly in small studies. Whether it becomes a useful clinical tool depends on larger validation studies and on how well patient privacy and consent are protected.
Sources
- Design and feasibility of smartphone-based digital phenotyping for long-term mental health monitoring in adolescents, Huang et al., PLOS Digital Health, July 2025
- Digital Phenotyping for Adolescent Mental Health: Feasibility Study Using Machine Learning to Predict Mental Health Risk From Active and Passive Smartphone Data, Kadirvelu et al., Journal of Medical Internet Research, February 2026 (arXiv preprint, January 2025)
- Multimodal Digital Phenotyping Study in Patients With Major Depressive Episodes and Healthy Controls (Mobile Monitoring of Mood), Aledavood et al., JMIR Mental Health, February 2025
- Predicting and Monitoring Symptoms in Patients Diagnosed With Depression Using Smartphone Data, Ikäheimonen et al., Journal of Medical Internet Research, December 2024
- One-Week Suicide Risk Prediction Using Real-Time Smartphone Monitoring, Journal of Medical Internet Research, 2023 (PubMed Central)
