Three New Clues in the Search for Depression Answers

An AI scanned thousands of wearables and flagged sleep chaos as a depression signal

What if the most reliable signal for depression severity is hiding in how erratically you fall asleep each night?

Here is the setup: you have wearable device readings from thousands of people — step counts, heart rate, sleep timing — and you want to know if any of those signals reliably track with depression severity. The CoDaS system, described in a new preprint, puts a multi-agent AI pipeline to work on exactly that problem. Think of it like a meticulous detective sitting inside a computer: it generates hypotheses about which signals might matter, runs the statistics, then cross-examines its own findings before passing them to a human expert for review. Across two independent depression datasets — DWB with 7,497 participants and GLOBEM with 704 participant-waves — CoDaS consistently found the same thing: sleep timing instability mattered. People whose sleep onset time varied a lot from night to night tended to score worse on depression measures. The correlation numbers are modest (ρ=0.126 to 0.252), meaning it's a real but weak signal, not a clean predictor. The team identified 41 candidate digital biomarkers for mental health overall. The catch is right there in the word 'candidate.' A correlation in an existing dataset is the beginning of scientific investigation, not the end. These signals have not been tested prospectively — meaning we have not yet followed real patients forward in time and asked whether wearable chaos predicts a future depressive episode rather than just reflecting one that's already there. Blinded human expert evaluators rated CoDaS's outputs highly compared to other systems, which is encouraging. But your smartwatch data is not a depression test yet. This paper is a map, not a destination.

A PNAS paper finally proposes a testable explanation for why ketamine lifts suicidal depression so fast

Ketamine can lift suicidal thoughts within hours — and medicine has been using it without knowing why.

About 30% of people with depression do not respond to standard antidepressants. For some of them, ketamine — originally developed as an anaesthetic and now approved in a modified form for depression — can reduce suicidal thinking within hours. Standard antidepressants take weeks, sometimes months, to do anything at all. That gap has puzzled researchers for years, because if you understood why ketamine works so fast, you might be able to engineer something safer and more targeted. A new paper published in the Proceedings of the National Academy of Sciences — one of the most selective scientific journals in the world — takes a direct swing at an explanation. The hypothesis, developed through synthesis of existing literature rather than new experiments, centres on synaptic and circuit plasticity: the idea that ketamine doesn't just nudge the brain's chemical balance momentarily, but triggers rapid physical changes in how neurons connect and communicate. Think of the brain's network of connections like a garden where certain paths have become overgrown and effectively closed off. Standard antidepressants are like gentle, slow watering. The hypothesis suggests ketamine is more like aggressive pruning and replanting: it forces rapid structural change. This is a proposal, not a proven mechanism. The paper's value is that it lays out what experiments would need to be run to test or disprove the idea — which is genuinely useful, because 'actionable' is right there in the title. If the plasticity explanation holds up, it could guide a generation of faster-acting depression treatments without ketamine's drawbacks: dissociation, restricted prescribing, potential misuse.

Combining two types of brain scan lets AI spot depression with 85% accuracy

Depression leaves no visible mark on a brain scan — unless you know how to combine two very different types of image.

When a doctor scans a brain to look for depression, they face a basic problem: there is no obvious thing to see. Depression is not a tumour or a lesion. A research team decided to combine two scan types and teach an AI to read both at once. The first type — structural MRI, or sMRI — is essentially a high-resolution photograph of the brain's physical shape and volume. The second — resting-state functional MRI, or rs-fMRI — is more like a time-lapse: it records which regions of the brain activate together while the person is just lying still, not doing anything. Together they are like having both a building's floorplan and a video of how people actually move through it — each one tells you something the other misses. The team's model, tested on the public REST-meta-MDD dataset, used a 'dual cross-attention' architecture: a type of AI design that looks at both scans simultaneously and allows each scan to help interpret what the other is showing. The result was 84.71% accuracy in classifying patients as having major depressive disorder versus not, compared to 65% to 82% for earlier methods on the same data. That is a meaningful technical step forward. The catch is that this is a benchmark comparison on a single dataset, not a clinic-ready diagnostic test. There is no mention of correcting for the fact that different MRI machines produce subtly different images, and the model has not been validated on genuinely new patients from different hospitals. Brain scanning is expensive and slow; making it diagnostically sharper is worth pursuing, but real-world deployment is still a long way off.

Look at what these three papers are actually trying to do and a pattern emerges: depression research right now is in an intense phase of instrument-building. We do not yet have a reliable blood test, a definitive brain scan, or a clear biological explanation for the fastest-acting treatment we have. What today's papers show is three different teams attacking that same deficit from three angles — passive sensing from your wrist, active imaging of your brain, and theoretical dissection of a drug that already works. None of them cross the finish line. The wearable signals are weak, the brain scan hasn't left the benchmark, and the ketamine paper is a hypothesis waiting for experiments. But here is what I'd argue: the gap between 'interesting correlation' and 'clinical tool' is exactly where mental health research has been stuck for decades, and these are precisely the kinds of incremental, unglamorous steps that eventually close it. The instruments are getting sharper. Slowly.

The most important near-term question from today's batch is whether CoDaS's wearable sleep signals hold up in a prospective study — one that follows people forward in time rather than looking backwards at existing data. That would be a materially different kind of evidence. On the ketamine front, watch for experimental work testing the synaptic plasticity hypothesis in the next year or two; if it replicates in animal models, it will attract serious drug development attention. No major conferences or trial readouts are imminent from these specific papers, but the wearable biomarker space moves fast — expect follow-up preprints within months.

• Wikipedia: What is a digital biomarker?
• Wikipedia: Ketamine — background and medical uses
• Wikipedia: Major depressive disorder — overview