Your sleep schedule, your stories, your voice: depression's new data trails

Your sleep's irregularity, not its length, may signal depression

Your smartwatch is already logging something that might matter for mental health — nobody just knew how to read it properly.

A research team built an AI pipeline called CoDaS that works like a very patient, very methodical plumber. Instead of checking the one tap you complained about, it maps pressure fluctuations across every pipe in the house — looking for patterns that suggest something is going wrong before the leak appears. In this case, the pipes are wearable sensors: steps, resting heart rate, sleep timing. The house is 9,279 participants across three real-world datasets. CoDaS runs in six stages: it profiles the raw data, generates hypotheses about which patterns might matter, tests them statistically, actively tries to break its own conclusions through adversarial checks, hunts for mechanistic explanations, and writes a report. Out of that process, it surfaced 41 candidate digital biomarkers for mental health. The two that held up across two separate depression cohorts were both about sleep irregularity — specifically, how much your sleep duration and your sleep onset time vary from night to night. Not whether you sleep too little. Whether your schedule is unstable. The correlation with depression symptoms was modest — a Spearman rank correlation around 0.13 to 0.25 — but it appeared consistently across different populations. Here is the catch, and it is a big one: 'candidate' is doing enormous work in that sentence. These are leads worth investigating, not clinical tools. A correlation of 0.25 means sleep variability explains roughly 6% of the variation in depression scores — real, but nowhere near diagnostic. Wearable datasets are also notorious for demographic skew and noise. This pipeline finds promising signals. The rigorous clinical trials that would validate them as actionable markers are a separate, longer project that has not started yet.

How you structure a story predicts your mental health better than your word choices

The rhythm and shape of how you tell a story — not the words you pick — turns out to be a stronger signal of mental health than any vocabulary test.

Eight hundred and thirty people wrote about difficult experiences as part of therapeutic writing programs in China — after disasters, in schools, in clinical settings. A research team then analysed those pieces of writing through three lenses simultaneously. The first looked at word choice: are there more negative words, more self-focused pronouns? The second looked at overall meaning, using semantic embeddings — a kind of dense mathematical fingerprint of what the text is about. The third looked at narrative structure: does the writer set a scene before explaining a crisis? Do they establish a clear before-and-after arc? Do ideas flow coherently from one to the next? Think of it like music. You can analyse a piece by cataloguing the individual notes (words), by capturing the overall melody (general meaning), or by studying the rhythm and structure of the whole performance — where the composer hesitates, rushes, or loses the thread. Turns out structure tells you the most. Two specific patterns emerged. Depression was linked to temporal disorganisation — writers jumped around in time, losing the thread of what came first. Anxiety was linked to spatial grounding deficits — writers struggled to anchor their stories in a concrete place or setting. These are computable, not just poetic. The catches: every writing sample here is Chinese-language therapeutic writing, which limits how far the finding travels into other cultures and languages. The large language model doing the narrative evaluation is also, frankly, a partial black box — we cannot be certain what structural features it is weighting most. And the full statistical breakdown was not available in the version the team published, so exact accuracy numbers remain pending peer review. Interesting enough to watch, not settled enough to act on.

An AI maps the web of a clinical conversation to detect depression

What if you mapped every statement in a clinical interview as a network — and let an AI navigate that network to spot depression?

Picture a detective who, after a long witness interview, does not just re-read the transcript. They draw a map: who mentioned what, in what order, which topics linked to which emotions, how the person's affect shifted across the conversation. That is roughly what PsyGAT does with structured clinical mental health interviews. A research team built a system that takes transcripts — where a clinician asks questions and a patient responds — and turns each session into a graph. Each statement becomes a node. The connections between statements, weighted by personality context extracted from the whole session, become edges. The AI navigates that web using a technique called a Graph Attention Network — essentially, it learns which connections matter most for predicting whether the person meets criteria for clinical depression. Tested on two standard benchmarks, DAIC-WoZ and E-DAIC (both datasets of real clinical interview transcripts), PsyGAT hit a Macro F1 score of 89.99 and 71.37 respectively — a measure of classification accuracy that corrects for the fact that depressed participants are always a minority in the data. The team reports outperforming several competing systems. Three catches. First: DAIC-WoZ has a very small training set — around 57 depressed participants. Numbers this high on small samples deserve healthy scepticism. Second: the paper claims to outperform 'GPT-5,' which is not publicly available for this benchmark task — exactly what was compared is unclear. Third: classifying patterns in a structured research dataset is not the same as helping a clinician diagnose a real patient sitting across from them. That gap remains wide and largely uncrossed.

Here is what strikes me about today: all three papers are chasing the same underlying question from different angles. Can we find reliable signals of depression in data that already exists — in your wrist sensor, in a piece of therapeutic writing, in a clinical transcript — before anyone has to explicitly name the condition? That shift is real and worth tracking. We are moving, slowly, from 'what do depressed people say?' to 'how do their patterns differ — in time, structure, and sequence?' But I want to name the shared limitation honestly. None of these are clinical tools. All three involve small or single-population samples, modest effect sizes, and no prospective validation — which means no study has yet followed people forward in time to check whether catching these signals early actually changes anything for them. The distance between 'we can detect a pattern in a dataset' and 'we can help a person earlier' is still the main problem in this entire field. Today does not close that gap. It makes the map a little more detailed.

The most important next step for CoDaS's wearable biomarker findings is a prospective trial — following people forward in time with their wearables and checking whether sleep variability actually predicts depression onset, not just correlates with it in cross-sectional data. No such trial has been announced yet. For the narrative work, replication outside Chinese-language therapeutic writing is the obvious test — it would be fascinating to see whether temporal disorganisation as a depression marker holds in English or Arabic storytelling contexts. Keep an eye on the DAIC-WoZ leaderboard too, since PsyGAT's claims invite a direct replication from independent teams.

• Wikipedia: Digital biomarker — background on what counts as a digital health signal
• Wikipedia: Computational psychiatry — the broader field these papers sit inside
• Wikipedia: Narrative therapy — context for why story structure matters clinically