Machine Learning Matched the Right Lifestyle Fix to Each Depressed Patient, Doubling Remission Rates

Fifty adults, mild-to-moderate depression, a Samsung smartwatch, and a smartphone app that pinged them four times a day. On the surface, the Personalized Mood Augmentation trial at UC San Diego sounds modest. What came out of it was not. Fifty-five percent of participants no longer met criteria for depression after just six weeks, roughly double the remission rate of standard clinical interventions. And the cognitive tests improved too, which nobody was quite expecting.

Depression affects about 21% of American adults at any given time. The economic burden sits somewhere north of $380 billion annually. The treatments available, drug and talk therapy alike, work for some people some of the time, but the average remission rate across clinical trials hovers around 30%. Something, clearly, is not adding up.

Jyoti Mishra, associate professor of psychiatry at UC San Diego and co-director of the Neural Engineering and Translation Labs, thinks she knows part of the problem. Generic advice, however evidence-based, tends to dissolve the moment it meets an individual human being in a depressed state. Sleep eight hours. Exercise 150 minutes a week. Eat a Mediterranean diet. The trouble is that depression is not a uniform illness, and neither, it turns out, are its lifestyle triggers. For one person, disrupted sleep might be the primary driver of low mood. For another, it’s social isolation. For a third, diet. A recommendation designed for everybody is, in a meaningful sense, designed for nobody.

Building a Model for One Person

The PerMA trial flipped that logic. Rather than asking what helps depressed people in general, Mishra’s team asked what predicts low mood in this specific person, based on their own data collected over time. It’s a concept statisticians call N-of-1 modelling. And it’s harder than it sounds.

During the trial’s first phase, participants wore a Samsung smartwatch continuously, tracking heart rate, step count, calories, and movement patterns. Simultaneously, they completed brief questionnaires through a smartphone app up to four times a day, rating their current mood and logging recent sleep quality, what they’d eaten, how much they’d moved, and whether they’d spoken with anyone. Sixty such sessions over two to four weeks. From this dense, personalised data stream, the team built a machine learning model tailored to each individual, identifying which lifestyle variables most reliably predicted that person’s depressed mood, rather than mood in depressed people generally. Model accuracy across participants averaged about 75%, which is, given the noise inherent in human behaviour data, rather impressive.

The output of each model was not a report full of correlations. It was something more actionable: a ranked list of an individual’s top mood predictors, visualised using a technique called Shapley statistics (borrowed from cooperative game theory, of all places), which breaks down exactly how much each variable is contributing to the prediction. Coaches, in this case medical students with eight hours of training, reviewed the rankings and assigned participants to one of four intervention domains: sleep, exercise, diet, or social connection. Seventeen of the 40 completers were pointed toward social connection. Thirteen toward exercise. Five each to sleep and diet.

“Our goal was to figure out the top lifestyle factors driving the depressed state, which would be different for different people, and to find out if by targeting those factors through personalized coaching, people would actually feel better,” Mishra said.

What Actually Changed

For six weeks, each participant met once weekly with their coach via video call, roughly 20 minutes a session, working through an intervention in whichever domain the algorithm had flagged. Some participants were following a cognitive behavioural therapy protocol for insomnia. Others were incrementally increasing the types of physical activity they already did. Others still were working on expanding social contact, or shifting diet toward patterns associated with improved mood. The content differed by person. The structure stayed consistent. As Mishra put it: “Every person in the trial was doing different behavioral therapies that are already well-established in the literature depending on their top predictive factor.” The personalisation was not in the therapy itself but in which therapy was selected, and why.

The daily smartphone check-ins continued throughout. And something the researchers found particularly striking: improvement in depressed mood tracked specifically with improvement in the targeted lifestyle domain, not with changes across the board. Participants’ off-target domains, the ones the algorithm hadn’t flagged, didn’t budge significantly. Which suggests the effect was not simply about engagement, or about the attentional benefit of having a health coach checking in once a week. It was specific.

Depression scores on the standard PHQ-9 scale dropped by a mean of 3.5 points, with a Cohen’s d of 0.89. That’s a large effect size by the conventions of clinical psychology. Anxiety fell 36%. Quality of life improved. Working memory, selective attention, and interference processing, cognitive domains that depression reliably degrades, all showed significant gains. “Clinical trials show that most current interventions only show about a 30% benefit on average in terms of depression remission,” Mishra observed; “here we see a near doubling of that due to targeting the top lifestyle predictive factors with data-driven personalized coaching.”

The Machine Could Do Most of This Itself

There is, buried in the paper, a finding that perhaps deserves more attention than the headline remission numbers. After the trial was complete, the researchers tested whether a large language model, specifically Google’s Gemini 2.5 flash, could replicate the iMAP assignments the human coaches had made from the same Shapley data. It matched 92.5% of the time. A simple rule-based decision algorithm, tuned empirically against the actual coach decisions, hit 95%. The human coaches were, in other words, doing something that could largely be automated. Which raises an obvious question about what happens when this approach gets scaled.

As Mishra noted, the generic advice we’re all given, healthier diets, more sleep, more exercise, is not wrong exactly; it’s just not especially useful when you’re depressed and can barely function. “When one is in a depressed state, it’s not possible to change everything about one’s life,” she said. “You’re just trying to survive and function on a day-to-day basis.” Personalized insights, the argument goes, are more tractable precisely because they’re narrower. Fix the one thing that matters most for you, rather than attempting an overhaul.

There are limits to what this trial can claim. Fifty participants, no control group, a single academic medical centre in San Diego. The benefits persisted at 12-week follow-up, but longer-term durability is unknown. And the effect size, impressive as it is, needs replication in a randomised controlled trial before anyone should revise clinical guidelines. What the PerMA trial offers is not a treatment. It’s a proof of concept: that a wearable, a smartphone, and a machine learning pipeline can identify which lever to pull for a given individual in ways that population-level research cannot. Whether that insight, translated into a scalable digital product, could eventually reach the two-thirds of depressed people who currently get inadequate care is still, for now, an open question.

https://doi.org/10.1038/s44277-026-00062-3

Frequently Asked Questions

Why didn’t researchers just tell everyone to do the same thing, like exercise or sleep more?

Because the lifestyle factor that most strongly predicts low mood varies considerably from person to person. In this trial, the machine learning models confirmed that social connection was the top driver for the largest group of participants, while exercise mattered most for others, and sleep or diet for others still. A blanket recommendation would be correct for some people and irrelevant for many more. The trial’s results specifically showed that mood improved in the targeted domain only, not across unrelated lifestyle areas, which suggests the personalisation was doing real work.

Could an app do this without a human coach?

Possibly, and sooner than you might think. The researchers tested whether a large language model could replicate the human coaches’ decisions about which lifestyle domain to target, and it matched in 92.5% of cases. A fine-tuned algorithmic approach hit 95%. The coaches’ role in interpreting the machine learning output appears to be largely automatable, though the researchers caution that some human oversight remains important for now, particularly for safety monitoring.

Is this approach suitable for people with severe depression?

Not as currently designed. The trial enrolled only people with mild-to-moderate depression, and deliberately excluded those with active substance use disorders, psychotic disorders, bipolar disorder, or acute suicidal behaviour. The intervention is built around lifestyle change, which requires a baseline level of functioning. For people with severe depression, it would likely need to be combined with other treatments rather than used as a standalone approach.

How long did the benefits last?

Improvements in depression scores were statistically significant at both the six-week and twelve-week follow-up points after the intervention ended, though sample sizes at those time points were smaller due to dropout. Anxiety improvements held at six weeks but were no longer statistically significant at twelve weeks. The researchers are clear that long-term durability data, and a proper randomised controlled trial, are needed before this moves from promising pilot to established treatment.