Machine learning made appropriate lifestyle modifications for each patient with depression, doubling remission rates

50 adults with mild to moderate depression, a Samsung smartwatch, and a smartphone app that pings four times a day. On the surface, the Personalized Mood Augmentation Trial at the University of California, 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, approximately twice the remission rate of standard clinical interventions. And cognitive tests improved, which no one expected.

Approximately 21% of American adults suffer from depression at any given time. The economic burden is more than $380 billion annually. Available treatments, such as medications and talk therapy, are sometimes effective for some people, but the average remission rate across clinical trials hovers around 30%. Clearly something isn’t adding up.

Jyoti Mishra, associate professor of psychiatry and co-director of the Institute for Neuroengineering and Translation at the University of California, San Diego, thinks she knows part of the problem. General advice, even when based on evidence, tends to disappear the moment you encounter an individual in a depressed state. Please sleep for 8 hours. Get 150 minutes of exercise a week. Eat a Mediterranean diet. The problem is that it turns out that depression is not a homogeneous disease and is not caused by lifestyle. For some people, disrupted sleep can be a major cause of low mood. Another is social isolation. The third thing is diet. Recommendations designed for everyone are not designed for anyone in any meaningful sense.

Building a model for one person

The PerMA trial overturned that logic. Rather than asking what helps people with depression in general, Mishra’s team asked what predicts low mood in this particular person, based on their own data collected over time. This is a concept that statisticians refer to as N-of-1 modeling. And it’s harder than it sounds.

In the first phase of the study, participants continuously wore a Samsung smartwatch that tracked their heart rate, steps, calories, and exercise patterns. At the same time, participants completed a simple questionnaire up to four times a day via a smartphone app to assess their current mood and record their recent sleep quality, what they ate, how much they moved, and whether they talked to anyone. Do 60 such sessions over 2-4 weeks. From this dense, personalized data stream, the team built a machine learning model tailored to each individual to identify which lifestyle variables most reliably predict a person’s depressed mood, rather than their mood in general. Model accuracy across participants averaged approximately 75%. This is quite impressive considering the noise inherent in human behavioral data.

The output of each model was not a correlation-filled report. It was more practical. It’s a ranked list of an individual’s top mood predictors, visualized using a technique called Shapley statistics (borrowed from cooperative game theory of all places) to break down exactly how much each variable contributes to the prediction. A coach, in this case a medical student who underwent eight hours of training, reviewed the rankings and assigned participants to one of four intervention areas: sleep, exercise, diet, and social connection. Of the 40 participants, 17 were aiming for social connections. 13 towards the movement. 5 for sleep and diet.

“Our goal was to understand the key lifestyle factors that cause depression, which may vary from person to person, and to see whether targeting those factors through personalized coaching could actually make people feel better,” Mishra said.

what actually changed

Over six weeks, each participant met with a coach via video call once a week and spent approximately 20 minutes per session working on interventions in areas flagged by the algorithm. Some participants followed a cognitive behavioral therapy protocol for insomnia. Some people gradually increase the type of physical activity they already do. Some were still working on expanding their social contacts and shifting their eating habits to patterns associated with improved mood. The content differed depending on the person. The structure remained consistent. As Mishra stated, “Each person in the trial received a different behavioral treatment that was already established in the literature depending on key predictors.” The individualization was not in the treatment itself, but in which treatment was chosen and why.

Daily smartphone check-ins continued throughout. And what the researchers found was particularly noteworthy: improvements in depressed mood specifically tracked with improvements in targeted lifestyle areas, rather than overall changes. Participants’ off-target domains, i.e., domains that the algorithm did not flag, did not change significantly. This suggests that the effect isn’t just about engagement or the attentional benefits of checking in with a health coach once a week. It was specific.

Depression scores on the standard PHQ-9 scale decreased by an average of 3.5 points, with a Cohen’s d of 0.89. This is a large effect size by the conventions of clinical psychology. Anxiety decreased by 36%. Quality of life has improved. Working memory, selective attention, and interference processing, cognitive areas that are reliably impaired by depression, all showed significant gains. “Clinical trials show that most current interventions are only about 30% effective on average in remitting depression,” Mishra said. “By targeting key lifestyle predictors with data-driven, personalized coaching, we’re seeing that number nearly double.”

A machine can do most things by itself

There are findings buried in this paper that are perhaps more noteworthy than the headline number of remissions. After the trial was completed, the researchers tested whether a large-scale language model, specifically Google’s Gemini 2.5 Flash, could reproduce the iMAP assignments made by human coaches from the same Shapley data. It matched 92.5% of the time. A simple rule-based decision algorithm empirically tuned to real coaches’ decisions reached 95%. In other words, the human coach was doing something that could largely be automated. This raises the obvious question of what happens when this approach is scaled up.

As Mishra pointed out, the common advice we’re all given – eat healthier, sleep more, exercise more – isn’t exactly wrong. It’s especially unhelpful if you’re depressed and can barely function. “When you’re depressed, it’s impossible to change everything in your life,” she said. “You’re just trying to survive and function day-to-day.” There’s an argument to be made that personalized insights are more manageable precisely because they’re narrower in scope. Instead of attempting a complete overhaul, fix the one thing that matters most to you.

There are limits to what can be argued in this trial. There were 50 participants, no control group, and one academic medical center in San Diego. The effect persisted at 12 weeks of follow-up, but long-term persistence is unknown. And the effect size, impressive as it is, needs to be replicated in a randomized controlled trial before anyone can revise clinical guidelines. What the PerMA trial provides is not a cure. This is a proof of concept. This means that wearables, smartphones, and machine learning pipelines can identify which levers to pull for specific individuals in a way that population-level studies cannot. Whether that insight can be translated into a scalable digital product that ultimately reaches the two-thirds of people with depression who are currently underserved remains an open question for now.

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

---

FAQ

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

That’s because the lifestyle factors that most strongly predict low mood vary considerably from person to person. In this trial, machine learning models confirmed that social connections were the biggest driver for the largest group of participants, while exercise was most important for others and sleep and diet remained important for others. Blanket recommendations may be correct for some people, but meaningless for many people. The results of this study specifically showed that mood improved in just the targeted area, rather than across unrelated lifestyle areas, suggesting that personalization really does work.

Can an app do this without a human coach?

Probably, and sooner than you think. The researchers tested whether a large language model could replicate human coaches’ decisions about which lifestyle areas to target, and found agreement in 92.5% of cases. Through a fine-tuned algorithmic approach, we reached 95%. Although the coach’s role in interpreting machine learning output appears to be largely automatable, researchers caution that human oversight remains important for now, especially when it comes to safety monitoring.

Is this approach suitable for patients with severe depression?

Not as currently designed. The trial enrolled only patients with mild to moderate depression and intentionally excluded those with active substance use disorders, psychotic disorders, bipolar disorder, or acute suicidal behavior. This intervention is built around lifestyle changes, which require a baseline level of functioning. For patients with severe depression, it may need to be combined with other treatments rather than used as a standalone approach.

How long did the benefit last?

Improvements in depression scores were statistically significant at both the 6- and 12-week follow-up time points after the end of the intervention, although sample sizes at these time points were small due to dropouts. Anxiety improvements were maintained at 6 weeks but were no longer statistically significant at 12 weeks. Researchers are clear that long-term durability data and appropriate randomized controlled trials are needed before this can move from a promising experimental treatment to an established treatment.

related