Three New Clues About What Shapes Your Mental Health

Your Fitness Tracker's Messy Sleep Schedule May Signal Depression

What if the early warning sign for depression isn't how you feel on Tuesday morning, but how inconsistently you've been going to bed all month?

A team building an AI system called CoDaS fed it wearable data from 9,279 people and asked it to hunt for hidden patterns that correlate with depression and metabolic disease. Think of it like hiring a very patient assistant to go through a year's worth of your fitness tracker logs — not looking for one bad night, but for the overall wobble in the rhythm. What CoDaS found, across two separate depression cohorts, was that people with depression tend to have more variable sleep timing: not just sleeping less, but sleeping at different hours each night. Sleep duration variability flagged depression in the larger cohort (7,497 people), and irregular sleep onset did the same in a second, independent group of 704 observations. In total, the system identified 41 candidate digital biomarkers — measurable signals from wearables — for mental health. Why does this matter? If these signals hold up, a fitness tracker you already own could flag early warning signs before someone even recognises they're struggling. That's a genuinely different world from the current one, which usually requires a clinic visit and a long conversation before anything is caught. The catch — and it's a real one — is that 'candidate' is doing heavy lifting in that sentence. These are correlations, not causes. The actual prediction improvement from adding CoDaS-discovered features was modest: an R² increase of about 0.04 for depression. That's a small but real step. The system was also tested on existing datasets, not a prospective clinical trial where you'd follow people forward in time. We don't yet know whether these signals are reliable enough to act on in a clinical setting. That test still has to happen.

Using AI as a Therapist May Hurt You. Using It as a Tool Might Not.

Same app, same chatbot, completely different effect on your brain — depending on whether you're using it to draft an email or to feel less alone.

A team of researchers gave 222 university students high-resolution brain MRIs and asked them how they use AI tools. Not how much — how. They split usage into functional (drafting, researching, organising) and socio-emotional (processing feelings, seeking companionship, emotional support). The results pulled in opposite directions. Students who used AI functionally showed slightly larger volumes in the dorsolateral prefrontal cortex — the region behind your forehead involved in planning and working memory — and better-connected hippocampal networks. They also had higher GPAs. Students who used AI for socio-emotional needs showed the reverse: lower grey matter volume in regions involved in social and emotional processing, and higher rates of depression and social anxiety. Here's the important context: only 7% of students fell into the heavy socio-emotional use category, compared to 83% for functional use. So the worrying pattern applies to a minority, not the average user. Think of it like the difference between using your phone as a map versus using it as your only social contact — same device, very different outcomes. The catch is a significant one: this is a cross-sectional study, meaning everyone was measured at one moment in time. We cannot tell from this data whether using AI emotionally causes worse mental health, or whether people who are already struggling are more likely to turn to AI for emotional support. Both explanations fit the data equally well. The sample is also 222 mostly healthy students at one university — a narrow group. Longitudinal work, following the same people over time, is what's needed next. This paper raises the question clearly. It doesn't answer it yet.

AI Claimed to Detect Depression by Reading the Doctor, Not the Patient

An AI model scored 98% accuracy at detecting depression from a clinical interview — but it had never read a single word the patient said.

Researchers tested AI models designed to detect depression from transcripts of structured clinical interviews across three datasets: ANDROIDS (116 sessions), DAIC-WOZ (189), and E-DAIC (275). The models performed well. Then the team tried something pointed: they trained the same models exclusively on what the interviewer said — removing all patient language entirely. On the ANDROIDS dataset, the interviewer-only model reached a macro-F1 score of 0.98 — essentially near-perfect. The patient-only model scored 0.79. The same pattern held across the other two datasets. What the AI had actually learned was the structure of the interview itself: which scripted questions tend to appear in a depression session, in what order, at what positions. It's a bit like a student who gets 95% on every test not because they understood the subject, but because they figured out that question 7 is always a trick question and question 12 always has option C. This matters beyond academic tidiness. Dozens of published papers have used these same datasets to report impressive depression-detection results. If the models found this shortcut, some of those results are measuring the wrong thing entirely. Deployed in a real clinical setting — where interviewers don't follow a rigid script — those models would likely collapse. The fix the authors propose is simple but humbling: train and evaluate only on patient turns. The scores drop. But they're honest. Honestly, this is exactly the kind of finding that the field needs more of — unglamorous, slightly uncomfortable, and genuinely useful.

These three papers are, in some ways, telling one story. We have more data about mental health than ever before — from wrists, from brain scans, from transcripts of clinical conversations. And we have more powerful tools to analyse it. CoDaS shows what wearable data could eventually do if its candidate signals survive prospective testing. The AI-use study shows that the technology we're already living with is shaping our mental health in ways that depend not on the tool itself but on how we reach for it. And the interviewer-bias study is a warning shot aimed directly at the field: some of the 'impressive' AI results in clinical research are less impressive than they look, because the models found a structural shortcut that won't survive contact with real clinical practice. The message isn't pessimistic. It's that the pipeline from a promising finding to a trustworthy clinical tool is considerably longer than the headline suggests — and papers like the third one are doing the unglamorous, necessary work of making that pipeline honest.

The interviewer-bias finding suggests that published depression-detection benchmarks should be systematically re-evaluated for this kind of leakage — watch for replication studies or challenges to the DAIC-WOZ and ANDROIDS datasets specifically. For CoDaS, the next real test is a prospective cohort study where wearable signals are collected before diagnosis, not retrospectively; that design hasn't appeared yet. And if the AI-use and brain structure finding interests you, a longitudinal version of that study — following the same students over two or three years — would be the paper worth waiting for.

• Wikipedia: Digital biomarkers — background on what they are and how they're validated
• Wikipedia: Interviewer effect — the broader social science problem the third paper is building on
• Wikipedia: Voxel-based morphometry — the brain imaging method behind the AI-use study