Changes in movement, sleep or speech can be associated with mental-health symptoms. Digital phenotyping studies those patterns using data from everyday devices. The useful question is what a particular signal measures, and how well it worked in the population tested.

What digital phenotyping means

The idea predates Thomas Insel's 2017 discussion of digital mental health. Jain, Powers, Hawkins and Brownstein published "The digital phenotype" in 2015. The term describes a way to study behavior through digital traces.

Researchers can use sensor data such as location and activity, or examine speech and interaction with a device. Surveys and check-ins add information that passive signals don't provide. A change in location patterns, for example, could reflect symptoms, a holiday or a new job; the sensor alone doesn't explain the reason.

In Barnett and colleagues' 2018 pilot, 17 patients with schizophrenia used the Beiwe app for up to three months. The rate of detected behavioral anomalies was 71% higher in the two weeks before relapse than during other periods. This is a relative increase in anomaly frequency, not 71% prediction accuracy. The study found an association preceding relapse; it didn't establish that a phone alerted a patient or doctor before either noticed symptoms. It also wasn't a depression study.

Platforms such as Beiwe make it possible to collect data outside appointments. In this type of monitoring, changes from a person's own routine can be worth investigating alongside absolute values. An individual baseline gives context, but it doesn't remove the need to check what caused a change.

What conversations reveal

Voice research examines features such as pitch variation, speech rate and pauses. Briganti and Lechien's 2025 systematic review included 12 studies and 16,872 participants. Reported AUC values for distinguishing depression from controls ranged from 0.71 to 0.93. Six studies had high risk of methodological bias. This is a range across study methods and populations, rather than one tool's expected performance for every speaker.

A concrete example is Mazur and colleagues' evaluation of Kintsugi Voice v1. It used English-language recordings of at least 25 seconds from adults in the United States and Canada. The validation set contained 4,456 participants. Against self-reported PHQ-9≥10, the reported sensitivity was 71.3% and specificity 73.5%; 920 validation samples received a separate "further evaluation recommended" output. The reference was a depression screening threshold, rather than an independent diagnosis, and these results don't validate anxiety detection.

Sensitivity describes how often a test identifies reference-positive cases; specificity describes how often it correctly excludes reference-negative cases. Neither number can be replaced with a generic "80% accuracy" claim.

Text research measures different things. Tauscher and colleagues analyzed 7,354 messages from 39 clients in a 12-week texting intervention. A fine-tuned BERT model identified messages with cognitive distortions at F1=0.62, compared with human-rater agreement of 0.63. The task was identifying distortions in these messages, not diagnosing depression or matching clinicians across all their work.

Researchers also combine audio, text and facial information. More channels can provide more context, but adding video doesn't automatically make a system more accurate; that needs a comparison on the same task and population.

Assessment inside a conversation

The MoPHES preprint describes two fine-tuned MiniCPM4-0.5B models, one for severity classification and another for dialogue. The authors explore mobile deployment so processing can happen on a device.

On their 200-sample classification benchmark, Table VI reports accuracy of 80.5% for anxiety severity and 63% for depression severity. The underlying text labels were generated using GPT-4o-mini rather than clinical diagnoses. MoPHES exceeded some larger models, while other models performed better on particular tasks. These figures describe label prediction on that benchmark, not diagnostic accuracy in a clinic.

The separation between assessment and dialogue is an engineering choice. It doesn't establish why one condition scored lower, or show that quietly assessing people without their knowledge would be an appropriate product feature.

Sensors and words answer different questions

Signal type	Example task	What to check
Location and activity	Detect a change from usual behavior	Personal baseline, alternative explanations and follow-up
Voice	Classify recordings against a screening threshold	Language, recording conditions, sensitivity and specificity
Text	Identify cognitive distortions in messages	Exact label, annotators and evaluation population
Conversation model	Predict severity labels in a benchmark	How labels were created and whether clinical validation exists

Sensor data can be continuous, while language provides information only when someone speaks or writes. Combining them creates a new system that needs its own evaluation. Our overview of computational models of mental disorders explains related modeling approaches; AI risk detection in text covers a different, higher-stakes prediction task.

Limits and practical questions

False alarms matter. If an app flags a change, the useful next question is what it expects the user or clinician to do with that information. A performance metric by itself doesn't establish that the resulting intervention helps.

Location history, voice and private messages also have different data implications. Ask what is collected, where it is processed and who can access it. Local inference describes one processing step; storage, backups and other services still depend on the product. We discuss those choices in on-device AI and privacy.

A screening or monitoring signal can prompt further assessment. Its interpretation depends on the task it was validated for. The guardrails article shows another example of why an evaluation method needs to stay attached to its result.

FAQ

Can a smartphone detect depression?

Researchers can use phone-derived signals to estimate associations with symptoms or screening scores. Performance varies by signal and study. A clinical diagnosis requires a different assessment from detecting a digital pattern.

Does AI read messages to assess mood?

Some research systems analyze message content, while others use activity or voice data. The product should explain which data it accesses and where the analysis runs. A research prototype's local-processing design doesn't establish another app's data policy.

How accurate is digital phenotyping?

There is no single accuracy figure. Kintsugi's reported sensitivity and specificity concern PHQ-9 screening, Tauscher's F1 concerns cognitive distortions, and MoPHES accuracy concerns benchmark labels. Comparing them as though they measure one task would be misleading.

Nearby offers emotional support and self-reflection, rather than diagnosis or treatment. The research above doesn't establish that Nearby uses these sensing methods; its facts page describes the product's actual scope.