Passive Sensors Are Watching. Is the AI Ready to Judge?

An AI watches cancer survivors' phones so they don't have to log moods

Cancer survivors often stop filling in mood diaries at the exact moment things get worst — so the PULSE team asked whether the phone could quietly keep watch instead.

Cancer survivors face elevated rates of depression and anxiety. But they also hit what the PULSE researchers call the 'diary paradox': self-reporting drops off sharpest exactly when distress peaks. You feel terrible, so the last thing you do is open an app and rate your feelings. The PULSE team's response: stop asking. Instead, let the phone observe passively — movement patterns, screen use, sleep timing, social communication — and have an AI agent piece together whether you're in a bad state and, crucially, whether right now is a good moment to send help. Think of it like a neighbor who notices you haven't collected your mail for three days, without waiting for you to call. The system isn't just a model sitting on raw numbers. It's an agent — meaning it runs its own mini-investigation, querying the phone data with purpose-built tools, making roughly five queries per check-in, in about 45 seconds. When tested on 50 cancer survivors, this approach hit a balanced accuracy of 0.743 for predicting 'does this person want to regulate their emotions right now?' — compared to 0.52–0.60 for traditional machine-learning approaches on similar tasks. Why does that gap matter? Because sending a mental health prompt at the wrong moment feels intrusive. Timing is almost as important as content. The catch: 50 people is a small room, not a stadium. The traditional ML comparison isn't a perfectly matched baseline — it's drawn from similar, not identical, experimental setups. And there's no safety framework yet for what happens when the system gets it wrong and sends — or withholds — support at the wrong moment. That question isn't answered here.

AI reads anxiety symptoms correctly — then decides you're fine because you have friends

The AI read the symptoms correctly, then decided the person was probably fine — because the person also had a support network.

A team built a benchmark of 555 real semi-structured mental health interviews, labeled by clinicians using standard diagnostic criteria (SCID — essentially the rulebook psychiatrists follow). Then they ran five AI language models — LLaMA 3, DeepSeek, GPT-4o Mini, GPT-4.1 Mini, and GPT-5 Mini — through zero-shot screening: no training on these cases, just a prompt and a diagnosis request. Performance swung wildly. Accuracy ranged from 0.49 — barely better than a coin flip — to 0.86. But even the best numbers looked shakier under a stricter lens: Matthews Correlation Coefficient, which adjusts for how rare the condition is in the data, ran from 0.16 to 0.38. That's modest at best. The most striking part wasn't the accuracy table. The researchers looked hard at cases the AI got wrong — specifically, cases it missed — for anxiety and PTSD. Again and again, the model had correctly identified symptoms in the transcript, then talked itself out of flagging them because the person also mentioned coping strategies, a functional job, or a good support network. It's like a doctor who spots a stress fracture on the scan, then clears you for the marathon because you walked into the clinic without limping. Those protective factors are clinically real — they matter in diagnosis. But here the AI was letting them override explicit symptom evidence in a way no trained clinician should. The catch: the rationale analysis — the part that revealed this coping-language pattern — was a post-hoc look at one model only (GPT-4.1 Mini). That's a clue, not a confirmed finding. And all models were tested zero-shot, not fine-tuned for clinical use.

A €40 DIY headband that tracks your sleep stages — sort of

You can now build a forehead-worn sleep monitor from 3D-printed parts and sports-bra fabric for less than a dinner for two.

The gold standard for measuring sleep — polysomnography, or PSG — requires an overnight stay in a lab, electrodes glued across your scalp and body, and a bill that can run into thousands. Consumer trackers like Oura or Whoop cost £250–400 and are black boxes you can't modify. The OSSMM team asked: what's the floor? Their answer: a 3D-printed headband, €37.80 in parts, with electrodes made from conductive strips cut from commercial fitness chest-strap fabric. It sits across your forehead and picks up electrical signals from your brain while you sleep. Think of it like rigging up a rooftop weather station versus subscribing to a commercial forecast service — less polished, fully open, and you can see every component. The signals feed a machine-learning model that classifies your sleep into four stages: awake, light sleep, deep sleep, and REM. Over 15 nights, the best model hit an accuracy of 0.776 and a Macro F1 score of 0.770 — numbers in the same rough territory as some consumer wearables. Sleep is a direct window into mental health: poor sleep is both a symptom and a driver of depression, anxiety, and bipolar episodes. A cheap, open-source device that a researcher in a low-resource setting — or a curious tinkerer — can actually build and modify has real value. But the catch here is significant: this was tested on exactly one person — one of the researchers — across 15 nights. The reference device they compared against wasn't even proper PSG; it was a consumer device that itself only agrees with lab results about 63% of the time. You're measuring one approximation against another. Real validation needs dozens of participants, diverse sleep conditions, and genuine polysomnography comparison.

Step back from the three papers and a pattern emerges: the field is building better sensors faster than it's building the judgment to interpret what those sensors find. PULSE wants your phone to silently track whether a cancer survivor needs support. OSSMM wants a €40 headband to map sleep as a proxy for mental state. And the LLM screening paper is the corrective that ties them together: even when the signal is clear and the symptom evidence is right there in the transcript, an AI can still talk itself out of acting on it because the surrounding context looked reassuring. That gap — between detecting a signal and knowing what to do with it — is where the real work lives. Cheaper hardware and smarter algorithms are necessary. They are not sufficient. The question of when 'the system sees something' should translate to 'a human gets help' is dangerously underspecified across all three papers. That's not a criticism. That's the problem to solve next.

The most important open question from this week: what happens when passive sensing systems get it wrong at scale — and who's liable? No paper here addresses it. On the tools side, keep an eye on whether the OSSMM team runs a multi-participant validation; the single-person limitation is the only thing standing between this project and a genuinely useful public resource. And if you're curious about the LLM screening work, the next step would be the same coping-language analysis applied across all five models — right now it's a hypothesis built on one.

• Wikipedia: Polysomnography — how clinical sleep testing actually works
• Wikipedia: Passive sensing in mobile health research
• Scientific American: The problem with AI in mental health screening