Digital Phenotyping: A Developer's Guide

There is a method for measuring human behavior that has a decade of peer-reviewed literature behind it, works on hardware most of the planet already owns, and has produced systematic reviews across depression, anxiety, bipolar disorder and schizophrenia.

Almost none of it has shipped in a product.

That gap is the subject of this article. Digital phenotyping is not a speculative technique waiting on better sensors. The sensors are in every pocket, the correlations replicate, and the models reach genuinely useful discrimination. It is stuck on things that never appear in an abstract: how much data you actually get, how fast it degrades, and whether any two teams compute the same feature the same way.

What digital phenotyping actually is

The term was introduced by Jukka-Pekka Onnela at the Harvard T.H. Chan School of Public Health, who defined it as the moment-by-moment quantification of the individual-level human phenotype in situ using data from personal digital devices [1].

A phenotype is an observable characteristic. Measuring a behavioral one traditionally meant asking: a PHQ-9 in a clinic, a sleep diary, a recall survey. That gives you a low-resolution snapshot filtered through memory and self-report bias, collected on the days someone happened to attend an appointment. Digital phenotyping replaces the snapshot with a continuous stream collected in situ, meaning in ordinary life rather than in a clinic.

The split that matters operationally:

• Active data requires effort. Surveys, symptom check-ins, tapping tests, voice samples. High signal, low volume, and it decays the moment engagement drops.
• Passive data requires nothing from the user. Accelerometer, location, screen state, ambient light. Sensors already running for other reasons.

Most studies combine both, and as the accuracy numbers below show, that combination is not a convenience. It is currently what makes the models work.

What the signals actually predict

Most explainers stop at “movement patterns correlate with mood.” The real numbers are more specific, and more useful.

The foundational work measured GPS-derived features against PHQ-9 scores, and the pattern has broadly held since [2]:

Read the signs. Every mobility feature is negatively correlated. Higher depressive symptom severity meant visiting fewer distinct places, distributing time less evenly across them, keeping a less regular daily rhythm, and staying home more. Phone use went the other way. Using normalized entropy alone, a classifier separated participants scoring PHQ-9 at or above 5 from those below it with 86.5% accuracy [2].

A separate ten-week study of college students replicated the direction at weaker magnitudes, r between -0.43 and -0.46, and surfaced a detail worth keeping if you ever compute these yourself: the relationships were stronger on weekend data than weekday data [3]. Weekday mobility is dominated by obligation. Weekend mobility is closer to choice, and choice is where the behavioral signal lives.

How well it performs today

Correlations are not predictions. On the prediction side, the field currently looks like this:

• Depression screening, combined models (passive streams fused with brief self-report microsurveys): AUC 0.77 to 0.83 [4]
• Relapse prediction in serious mental illness, one to four weeks ahead: AUC 0.70 to 0.88 [5]
• Mood state prediction: accuracy 0.61 to 0.67, AUC 0.56 to 0.69 [5]

Three things follow.

Combined beats passive-only, consistently. The 2026 Korean community study behind those depression figures found combined models outperformed single-modality approaches for both depression and anxiety [4]. A few microsurvey taps a week buy real discrimination that sensors alone do not. If your product design assumes zero user input, you are choosing the weaker model on purpose.

Mood is much harder than depression. An AUC of 0.56 is barely better than a coin flip. Screening for a persistent state works considerably better than inferring a transient one, which makes intuitive sense: a two-week depressive episode reshapes your mobility footprint, a bad Tuesday does not.

These are screening numbers. An AUC of 0.8 is genuinely useful for deciding who to reach out to, and genuinely insufficient for telling someone they are depressed. Products built on this should be designed to prompt, not to pronounce.

The tooling exists, and it is not built for you

One reason this field looks stalled from the outside is that its infrastructure is real but pointed somewhere else. There are several mature, actively maintained open-source platforms, and they are good:

Every one of these is built for a study: a consented cohort, a fixed window, a research protocol, an analyst who can inspect and repair the data afterwards. None is built for an unbounded consumer population where data arrives forever, permissions change mid-stream, and nobody is available to patch a gap by hand. That is not a criticism of the tools. It is a description of what they were designed to do, and it is exactly the gap between a research pipeline and a product one.

The deployments are also broader than the funding map suggests. Beyond the Korean community study, protocols now run digital phenotyping to predict relapse in serious mental disorders among slum residents in Dhaka [16], a context where wrist-worn hardware is not a realistic assumption for anyone. The method’s reach is not limited to well-resourced settings, which is much of why it matters.

The part that decides whether any of it works

Here is what the abstracts leave out.

Missingness is enormous and rarely reported. Across digital phenotyping studies monitoring psychosis and schizophrenia, missingness ranged from 19% to 72% [6]. That is not a spread you can average over. At the top of it, most of the behavioral record for that cohort does not exist.

Passive data is not free of engagement. This is the misconception worth correcting, and I have made it myself. GPS and accelerometer data quality can decline precipitously within about three days if a participant stops opening the app [7]. Both mobile operating systems throttle background execution for apps the user is not actively engaging with, so “passive” describes the user’s effort, not the collection’s independence. Continuous background collection is something you engineer against the OS, not something you are given.

The hardware fights you too. A comparative assessment of Android and iOS collection in real-world conditions found data integrity at 96% on devices whose power-saving policies did not interfere with sensor management, and 84% overall once aggressive manufacturer power management was included [8]. Continuous sensing also carries a measurable battery cost, roughly 12% at moderate sampling rates [8], and battery drain is itself a cause of uninstalls.

The ground shifts under you. Android ships on the order of 100 API updates a month [9]. A sampling strategy that was reliable last quarter can quietly stop firing.

The standard mitigation is real and worth knowing: planned duty cycling, meaning predefined on-off sampling schedules that keep battery draw low enough that the operating system does not start restricting you [9]. That is an engineering discipline, not a configuration flag.

Cohort attrition follows the same pattern. The 2026 Korean study enrolled 455 community-dwelling adults and, after applying predefined data-coverage criteria, retained 277 for the depression analysis [4]. Roughly two in five participants did not produce enough usable data to be analyzed, in a funded study with active participant management. A consumer app will not do better.

Why the field cannot yet agree with itself

The deeper problem is not any single study. It is that the studies cannot be compared.

A 2026 scoping review mapping smartphone-based digital phenotyping across health conditions found recurring methodological gaps: incomplete or inconsistent sensor descriptions, limited reporting of sampling rates and missingness, and heterogeneous validation practices [6]. A review of the depression literature met heterogeneity so severe that it precluded quantitative synthesis altogether [10]. The papers could not be meta-analyzed because no two of them computed “time at home” the same way.

This is why the field has an evidence base and not yet a standard. Two teams can both report that location entropy predicts depression, using different sampling rates, different window lengths, different definitions of a “place,” and different handling of missing days, and produce results nobody can pool.

The practical translation for anyone building on this is blunt: you cannot lift a feature definition from a paper and expect the reported performance. That number came from that paper’s pipeline, on that paper’s cohort, with that paper’s missingness. The feature is portable. The performance is not.

Where the regulators are

The 2026 position is more permissive than it was, and still short of validation.

On 6 January 2026 the FDA updated its guidance on Clinical Decision Support Software and General Wellness products, a meaningful shift toward a less restrictive posture for consumer wearables and AI-driven tools [11]. On 31 March 2026 it issued a request for information on the use of digital health technologies in clinical investigations for drugs and biologics [11].

But no digital biomarker in this space has achieved qualification as an FDA drug development tool or at the EMA equivalent [12]. Digital phenotyping has moved from academic curiosity to embedded feasibility work inside depression, bipolar and schizophrenia trials. Converting a measurable signal into a clinically actionable one is the bar most published work has not cleared [12].

So: build wellness and engagement products on this today. Do not build clinical claims on it.

What this means for builders

The phone is not the consolation prize. The literature did not settle on smartphones because wearables were unavailable. It settled on them because the signals carrying the most predictive weight, meaning mobility, routine, circadian regularity and interaction patterns, are phone signals. A wearable adds physiology on top. It does not replace the behavioral layer, and for most of your users it will never be there at all.

Budget for missing data as a first-class problem, not an edge case. Between 19% and 72% missingness, three-day quality decay, 84% device integrity and 40% cohort attrition in a managed study, the realistic assumption is that a large share of your behavioral record will be absent, and absent non-randomly. Every downstream feature needs an explicit answer for what it does when its inputs are partial.

Do not skip the microsurvey. The gap between combined and passive-only models is the difference between a product that works and one that technically runs. Two taps a week is a small ask for the discrimination it buys.

Consistency is the scarce asset, not sensor access. Anyone can read an accelerometer. What the literature demonstrably lacks is the same feature, computed the same way, across a large and uncontrolled population, with missingness handled identically every time. That is the layer we build at Sahha, and the strategic point holds whether you build it or buy it: the papers are public and reproducible in principle, and the reason so little has shipped is that the engineering underneath them is where the years go.

Ten years of evidence, and the bottleneck was never the science.