Your watch, your voice, and the AI that fakes patients

Your Fitbit's sleep data might be a depression signal

What if the messiness of your sleep schedule — not just how long you sleep, but how chaotic the timing is — turned out to be one of the clearest signals that depression is getting worse?

A team of researchers built CoDaS, a multi-agent AI system that acts like an automated lab assistant: it takes wearable sensor data, generates hypotheses about which patterns might matter, tests them statistically, then tries to poke holes in its own conclusions before writing a report. Think of it like a chef who cooks the dish, tastes it, invites a critic to trash it, and only then puts it on the menu. They ran CoDaS across three real datasets totaling over 9,000 participant-observations. The clearest finding: it wasn't how long people slept that tracked with depression severity — it was how erratic their sleep schedule was from night to night. Sleep duration variability correlated with depression scores in one cohort (ρ=0.252), and sleep onset variability — basically, whether you go to bed at wildly different times each night — correlated in a second, independent cohort (ρ=0.126). That replication across two separate datasets is the interesting part. The system also flagged 41 candidate biomarkers for mental health overall, though the word 'candidate' is doing heavy lifting there — a biomarker is only as useful as the clinical action you can attach to it, and that work hasn't started yet. The catch: those correlations are real but small. The depression prediction improvement over a basic demographic baseline was ΔR²=0.040 — a modest lift. CoDaS is a discovery tool, not a diagnosis machine. What it's doing is narrowing the list of variables worth studying properly in randomized trials. That's genuinely useful, but it's a very early step on a long road.

Depression leaves a fingerprint in how your voice wobbles

Not what you say, not even the tone of your voice — but the hidden rhythm of how your voice moves through sound-space during a conversation might mark you as depressed.

Researchers working with the DAIC-WOZ dataset — 142 people (100 without depression, 42 with) having structured clinical interviews — tried something unusual. Instead of measuring static properties of voice like average pitch or speaking rate, they tracked how a person's voice moves through a high-dimensional acoustic space moment by moment, then looked for a property called recurrence: how often does the voice return to the same neighbourhood of that space? The intuition is like watching someone pace a room. A non-depressed person might wander freely. A depressed person might get stuck in the same corner, returning again and again. Depression, this paper argues, changes the 'pacing pattern' of the voice in a measurable way. Using a technique called recurrence quantification analysis — essentially counting how often vocal trajectories loop back on themselves — the team achieved a cross-validated AUC of 0.689. That means, on average, the model correctly ranks a depressed person above a non-depressed person about 69% of the time, compared to the 50% you'd get by flipping a coin. Better than the baselines they tested? Yes. Clinically useful on its own? Not yet. The bootstrap confidence interval on their pooled AUC runs from 0.568 to 0.758 — that's a wide range, reflecting real uncertainty from a small sample. With only 42 depressed participants, every fold of the cross-validation is fragile. What this paper gives you is a proof of concept: the signal is probably there. Finding it reliably will take much larger studies.

AI can fake a convincing patient but lie about the whole population

AI models can roleplay a depressed patient so convincingly that no clinical rule is violated — and yet, at the population level, they're fabricating a reality that doesn't exist.

Researchers generated 28,800 synthetic patient profiles using four major AI models — GPT-4o-mini, DeepSeek-V3, Gemini-3-Flash, and GLM-4.7 — asking each to simulate patients from 120 different demographic combinations (varying race, gender, socioeconomic status, relationship status). They then checked whether the simulated population matched real epidemiological data from large U.S. health surveys. Here's the uncomfortable finding. Every single simulated patient followed DSM-5 clinical rules perfectly — zero violations across 28,714 cases. Ask one AI patient a question and the answers hold together. But zoom out to the whole crowd, and the picture is wrong. Think of it like a costume shop that makes a perfect outfit for any one person who walks in, but if you looked at everyone wearing their costumes at once, somehow every costume is a medium — the extra-smalls and extra-larges have disappeared. That's what happened here. The models compressed the real diversity of depression severity by 14% to 62% depending on the model, erasing the extremes of clinical reality. They also systematically overestimated depression severity for most groups by 3.6 to 6.1 PHQ-9 points — a substantial drift — while simultaneously underestimating depression in transgender women by over 5 points, capturing only a fraction of documented minority stress. Why does this matter? Because AI-simulated patients are already being used to train clinical tools and test therapeutic chatbots. If the training population is quietly wrong, the tools built on it will be wrong in ways that are hard to see. This paper doesn't offer a fix yet. But it makes the problem impossible to ignore.

Here is what today's three papers collectively say, if you read them in order: we are getting increasingly good at finding traces of depression in places we didn't look before — the chaos in your sleep schedule, the looping rhythm of your voice. Both are real signals, both are modest in size, both need much larger replication. That's the honest state of biomarker discovery: early, promising, fragile. But then comes PsychBench, which should make you cautious about the infrastructure being built around these discoveries. AI models that seem to perform well on individuals can silently fail at the population level — erasing clinical extremes, misrepresenting minority groups, looking trustworthy on every metric until someone thinks to check the right one. The field is simultaneously making real progress on detection and building evaluation systems that may not catch the failures that matter most. That tension is the thing to watch.

The most important next step for the vocal biomarker work is independent replication on a larger, more diverse sample — the DAIC-WOZ dataset has been used so many times that it carries its own biases. For PsychBench, the question is whether model developers respond with concrete fixes or whether this becomes another finding that gets cited and ignored. Worth watching: any trial results from the wearable depression monitoring space in the next few months, where the gap between lab accuracy and real-world clinical utility tends to become brutally apparent.

• Wikipedia: What is a biomarker?
• Wikipedia: Circadian rhythm — why sleep timing matters
• Wikipedia: PHQ-9 depression scale explained