Can Your Phone Predict a Breakdown Before You Can?

AI Reads Your Phone Behavior and Predicts Anxiety Across Different Studies

Your phone already knows when your sleep went sideways three nights in a row — the question is whether an AI trained on someone else's phone can figure out what that means for you.

The usual problem with AI mental health tools: you train them on one group of people, they work reasonably well, then you try them on a different group — different country, different study design, different phone habits — and accuracy falls apart. Like a recipe that only works in the kitchen where it was developed. Great for the author, useless everywhere else. The team behind a system called TimeSRL tackled this with a two-step trick. First, instead of feeding raw numbers — step counts, screen-on time, sleep logs — directly into a prediction model, they translate that data into plain English first. Something like: "low physical activity over three days, irregular sleep onset, high phone use after midnight." Then a language model reasons from those plain-English summaries to a mental health score, using a standardized depression-and-anxiety questionnaire called the PHQ-4 as the target. The translation step is the real insight. Numbers vary wildly across devices and studies. Plain-language summaries capture the underlying pattern in a form that travels better across contexts. They tested this with a protocol called leave-one-study-out: train the model on all datasets except one, then test on that hidden one entirely. No peeking. In those tests, TimeSRL reduced prediction error for anxiety by 3–10% over conventional machine-learning tools, and by 9–44% over other AI language models. Depression improvements were similar. The catch: all the datasets involved are predominantly university students in high-income countries. Whether this generalises to a middle-aged nurse or a rural teenager is genuinely unknown. And a 10% error reduction sounds meaningful but doesn't yet tell us whether it would change what a clinician actually does. That step hasn't been tested.

AI Psychiatric Screeners Miss Cases When Patients Say They Are Coping

If someone describes classic anxiety symptoms but also tells you their friends are supportive and they're managing fine, should an AI flag them for clinical screening?

That's exactly the judgment call a research team put to five AI language models, using transcripts from 555 real people who had completed proper clinical diagnostic evaluations alongside in-depth interviews. The models — including versions of GPT-4o, GPT-4.1 Mini, GPT-5 Mini, LLaMA 3, and DeepSeek — were asked to read each interview and say whether the person had anxiety, depression, PTSD, or any mental health condition. Overall accuracy ranged from 49% to 86% depending on the condition and the model. That lower end is barely above a coin flip. But the more damning number is the Matthews correlation coefficient — a stricter metric that adjusts for how often a model gets lucky on imbalanced data, think of it like a score that penalises you for guessing the common answer. Across all models and all conditions, this score ranged only from 0.16 to 0.38. Clinicians generally want to see above 0.5 before trusting a tool in practice. The analysis of why models missed cases is the most useful part. When someone described clear anxiety or PTSD symptoms alongside language about coping well, having social support, or functioning okay at work, the AI consistently backed away from a positive screening result. Like a doctor who hears "chest pain" but then hears "I feel fine most days" and doesn't order the test — sometimes that's good clinical judgement, sometimes it's a dangerous miss. The catch: these models were tested without any task-specific training — zero examples, just instructions. A model fine-tuned on clinical data might handle protective-context language better. But the finding itself is a real warning for product teams building mental health features on top of general-purpose AI.

An AI Agent Tracks Cancer Survivors' Mental Health When They Go Silent

The people who most need a mood check-in are the ones least likely to open the app on their worst days.

The researchers behind PULSE, working with 50 cancer survivors, named this the "diary paradox": self-report tools fail precisely when they'd matter most. Their proposed fix is to lean on what phones passively record — GPS patterns, app usage, call frequency, screen interactions — rather than waiting for someone to actively log how they're feeling. But raw phone signals are messy. So the PULSE team built an AI that approaches each person less like a calculator and more like a detective. Instead of running fixed formulas — "take step count, multiply by weight, add sleep score" — the system can ask questions of the data across multiple steps. "What did the GPS pattern look like the two nights before the mood drop?" "Was screen use unusually high or low that week?" The system has eight purpose-built tools for querying different signal types, and it uses them in sequence, updating its reasoning as it goes. This iterative, question-asking approach is what's called an "agentic" architecture. The result: for predicting whether a survivor was in a state where they'd actually want an emotional support intervention, the agentic system hit 74% balanced accuracy. Traditional machine-learning baselines on similar passive sensing tasks tend to sit around 52–60%. The catch: 50 participants is a small proof-of-concept, not a clinical trial. The evaluation was also retrospective — the AI looked back at data where outcomes were already known. Real deployment, where it has to act before the moment passes, is harder. And the "does this person want an intervention right now" question is still a long way from "does an intervention actually help." Worth watching, not yet ready to deploy.

Three papers, one shared theme: AI tools for mental health are getting technically more sophisticated while the real-world validation trails further behind. TimeSRL shows that translating raw phone data into language first helps models generalise across populations — a real step toward tools that don't have to be retrained from scratch for every new clinic or cohort. PULSE shows that letting AI reason dynamically, rather than apply fixed formulas, closes a meaningful gap in prediction. But then the LLM screener study lands and asks: if these models can't reliably distinguish "this person has anxiety" from "this person has anxiety but seems okay today," what exactly are we building toward? The three together suggest the field is advancing the machinery faster than it's stress-testing the judgement embedded in that machinery. That asymmetry is the thing to watch.

The most important open question across all three papers is external clinical validation — do these accuracy gains survive contact with real healthcare settings, real patient populations, and real clinicians making real decisions? None of these papers answer that yet. I'd also watch for whether the LLM screener findings prompt systematic work on fine-tuning models specifically on clinical interview data — right now, the best-performing model still topped out at an MCC of 0.38, which is not a number anyone should be comfortable deploying at scale.

• Wikipedia: Patient Health Questionnaire (PHQ) — background on the depression/anxiety scales used in these papers
• Wikipedia: Ecological momentary assessment — the methodology behind real-time mood tracking that passive sensing is trying to replace