NEATCast Episode 6 - Personalized AI/ML guided Lifestyle intervention for Depression

Added: 2026-06-01

[00:00:00]Hi everyone, welcome to the Neatcast podcast episode 6. Uh this is the podcast where we talk about mental health advances applicable to everyone's daily lives. My name is Dr. Luxshan Ramen. I'm a psychiatrist and neuroscientist at UC San Diego. Um and you guys can introduce yourselves.

[00:00:20]>> Hi, I'm Dr. Joti Mishra. I'm a neuroscientist in the department of psychiatry at UC San Diego.

[00:00:29]Hi, I'm Dr. Jason Nan and I have my PhD in bioengineering but now I'm working as a posttock in psychiatry.

[00:00:37]So I'm very excited today to talk about a paper that we recently published. Um, this describes a personalized AI machine learning guided lifestyle intervention for depression, which is a mouthful, but actually is in my opinion the direction that we need to go in psychiatry. Um, so I'm just going to speak a little bit to that. So, you know, in psychiatry, we we have a for many treatments a one-sizefits-all kind of approach, right? So, um, when somebody comes to my office and they tell me they're depressed, they're anxious, I say, "Well, here's what the evidence says can be helpful."

[00:01:23]That evidence can say medications can be helpful. Psychotherapy can be helpful, but there's a lot of evidence that says lifestylebased treatments for depression can also be helpful. Right? These are things we often will tell people, you know, you need to look at what you're eating and eat healthy foods. You should exercise. These have been shown to help people with depression, anxiety, um mindfulness interventions, getting better sleep.

[00:01:51]A lot of times patients will say, "I know, doc. I know, doc. I should be doing these things, but which one is the most important?" And I'll say, "Well, I'm not sure which one is the most important." And that kind of gets into the fact that for different people, it's probably the case that different factors are important for them. And as a clinician, uh it's often a matter of guess and check, guess try trial and error, right?

[00:02:25]Say, well, why don't you why don't you try exercising for a little bit of time and see how that goes? And that's not very satisfying as a clinician, right?

[00:02:34]to not really know what to what to tell people or how to guide them and it's often not satisfying and it's overwhelming for patients. They'll say, "Oh, you're telling me to eat better and sleep better and like that's too much."

[00:02:46]Um, so this leads to this particular trial that Dr. Mishra spearheaded. Uh, I got to be part of it and associated with it. Um, so I wonder if you could speak a little bit to your own interests and motivations in this and tell us more about the paper and how how it worked and what you did.

[00:03:10]>> Sure, Dr. um, great to be here. We're all part of the Neat Labs team at uh, UCSD and uh, this project um, began um, actually 5 years ago or so. Um and I've really you know uh trained from a translational point of view of uh applying neuroscientific and neurochnology advances to mental health care and psychiatry and um my goals were really that of course we understand that psychiatric disorders are very complex and how can we leverage uh data especially scalable data that is lowcost and available in um many settings to understand the relationships between um brain, body and behavior and illness.

[00:04:00]And so um the motivations of why this project started were really to figure out how we can use um uh data from um subjective information which is from surveys that people can fill out on a daily basis about um how they're feeling and what they're doing with their lifestyles. like Duction mentioned, important factors that doctors consider uh related to sleep, exercise, diet, um connecting with others. And um in our first iteration of the study, we also had um uh EEG brain recordings as part of our modeling work that we first published in 2021. Um and uh our goals were really to say okay well if a given person who may come to us with depression uh has fluctuating uh mood over time how can we um understand all the different uh lifestyle factors that we collect from for from them from these uh different tools um from their smartphone from their from wearables from EEG which was not part of this specific study. But previously we've also integrated EEG. Um how can we use all of these different tools to really understand um what is the um you know the primary factors that are driving their mood on a day-to-day basis and maybe that can eventually lead to a more informed uh and personalized treatment.

[00:05:35]And um 5 years ago the goals were uh really to just come up with um analytics that would be at the personalized level.

[00:05:46]Um that in itself was an advance because most studies are done with the individual being understood as part of a group. Um and most studies they you know they will put um healthy individuals into a bucket and then people with depression or another disorder into another bucket and just compare groups where in this case our goal was to really say hey for a given individual who's very unique that comes to us what are their factors unique factors and um um just >> so I'm going to I'm going to interject real quick. So um what you did is to say people with depression their mood doesn't stay always depressed and I think that's an important point their mood everybody's mood fluctuates it gets better it gets worse and you try to collect information from them subjective information which is things that they can report and then other kinds of information to try to predict what drove somebody's mood, what made it better, what made it worse.

[00:06:54]>> Exactly. Um that was the the point because we wanted to understand that um for every individual there is uh this um you know there's not a strict point in on the scale that there's just one score of depression that we're looking at. um every day uh one can go through um better mood states versus wor worse mood states and uh no matter how depressed one might be feeling there might be times when you're feeling just slightly better than not and so if we can consider that sort of as a continuum then when you're feeling better what is exactly happening in your life that's making it better and um especially factors that are actionable and control and and could be in your control if you had insight on them. If you know um for some person it might be more exercise that they're more sensitive to, another person may eat more vegetables and that's when they feel better. And and given that we do so many things in our day, we may not be able to systematically track it ourselves, but um that's the power of engineering and data that can um provide these kinds of insights. And so um at Neat Labs um the the Ne in Neat Labs is really engineering and this part really would not have been possible without having um brilliant engineers um part of our team.

[00:08:24]And so Jason has spearheaded this engineering effort um since our first paper in 2021 and then beyond that. And so, um, Jason, uh, would love to hear from you if you can tell people about, um, what kind of, um, engineering went into this project, what you found challenging, and what you find, um, exciting in here.

[00:08:50]>> Yeah. So, um, well, the engineering aspect is like really exciting like J mentioned. And one thing that you know we kind of wanted to build around is this idea of passive sensing right every single day you know we're outputting as humans just infinite data one can think of and you know consumer hardware has also caught on to this fact and then if you look at five years ago till now if you look at the number of people wearing smart watches or Apple watches and and now aura rings or the whoop you know it seems like almost everyone Everyone is just collecting all of this data all the time. But you know, just by looking at it, people don't really know what it means. Like you you look at sleep. Okay, I slept really good yesterday. So what?

[00:09:41]You know, like, oh, my heart rate is between like 100 and 120 BPM resting, you know, but what does that really mean? Uh, and so it kind of gets really complicated to tease out all of these what do they means? And with you know the advances of like machine learning and AI uh one thing that it's really good at is being able to sort through and understand large amounts of data or big data. Uh you can think of it as. And so just you know like being able to integrate all these different consumer hardwares and seeing you know iteration after iteration of new hardware and being able to you know just keep adding on to this suite that we have just makes it way more robust to just anyone who wants to use it you know and then additionally you know building part of this platform the idea is that we want people to actually use it and how do we get people to use it? Well, you have to reduce the barrier of entry as much as you can. And so I'm happy to report that right now we have gotten to a stage where all you need to do is download an app, you know, record some data and link up whatever wearable you have. And from there, a machine learning model will be just automatically built through the app through a cloud server that we have. and you get a full print out of results and just seeing like seeing this after we publish the paper we have already gotten some people who are interested you know that they've emailed be like hey I really want to I really want to use this and it's very rare that you have a publication where you propose an intervention and it goes directly into the hands of the general public like within a matter of days of the publication coming out. So that was you know that that was really cool.

[00:11:40]>> Um and then >> I think there was two parts maybe you could speak to. One one part of the study was the machine learning model >> and then the second part of it was how did you and Dr. Misher you can speak to this. How did you take the model and then create an intervention out of it?

[00:12:03]>> Oh yeah. Yeah definitely. So um the model you know the the crux of it we we collect all these passive sensing data right and then we want some way to relate that to how someone is feeling.

[00:12:15]So their depressed mood in this case. So you know the first step is if we can build a machine learning model to predict someone's mood based on all their lifestyle attributes whatever that may be. You know the the idea is that if the model can accurately predict your depression, it at some level fundamentally understands what is causing you know one person's depression.

[00:12:38]And from there uh we can go one step further and that is to actually use something called model explainers and we can open up the machine learning model and see how it's making the determination. So we can start to understand why did it say this person would have you know high depression based on all this data point and if once we have that information right and this information comes in a sort of just list right most important features or most important variables that predict depression we can start looking at trends within that list. So for example, if we see a bunch of, you know, diet related variables as very important, then we say, oh, maybe this person needs a diet-based intervention. Conversely, if we see a bunch of exercise variables, an exercise intervention might be more suitable uh for them. And that in of itself is very powerful uh because going back to the idea of people see a bunch of data and say, okay, what does that mean? Sometimes we're very biased within ourselves in trying to relate what we're doing to how we feel. You know, one person might think that, you know, they exercise a lot and that their exercise doesn't trigger their depression or or whatnot and then so they start focusing on their diet. But it could be that you you know their understanding about how much they're exercising is fundamentally flawed or not tracking properly. So they have this internal bias and that could lead people down you know a fruitless endeavor. Um or you know if you have friends and they tell you hey I started doing this I started like walking 20,000 steps a day and I start feeling great.

[00:14:28]Then you might say, "Oh, well, if it worked for them, it's going to work for me." And so you do 20,000 steps a day, but you at the end of the day, you just feel very tired and sore and it actually makes your like depression worse, but then you keep going at it because your friend told you to do it.

[00:14:46]>> Yeah. And so that's >> and where we are.

[00:14:49]>> Yeah. Absolutely. I think um that's kind of the um really remarkable nature of um how we um came up with these individualized mood augmentation plans that we call in our paper IMAPs for short. And um Duction, I remember in our initial conversations um that was something that was really surprising to you that everybody did have a unique um uh footprint or an IMAP that was um that defined them. Um at least for the time that we were monitoring them. And keep in mind that this would change, you know, at different points in life. One could have a different IMAP than, you know, when you were being monitored um you know uh last year versus this year. And to Jason's point, um, you know, we've seen that in the data many times that, um, uh, it's really, um, potentially some people may have may be doing better when they're doing consistent low-level activity rather than more strenuous activity and or some people actually do better when they're um, you know, chatting with people individually and not having a big group gatherings too often. and um things that may seem counterintuitive, but that may promote an individual's well-being in a way that is um um that's insightful that they could, you know, keep better track of and um you know, promote and and uh help optimize um their well-being. And eventually it's kind of a closed loop system potentially, you know, for one person who's sleeping well 8 hours a day and they're um, you know, their their sleep quality is really good and um, they get to bed easily, then sleep is not to be optimized and it won't be surprising that that doesn't come up as a predictor of their fluctuating mood because it's pretty flat most of the time. uh but there are other inconsistencies that could be worked out and that are re relevant to their mood.

[00:16:54]Um so and also how Jason mentioned another thing that we're excited about this work is that it is um very immediately translatable to to a wider audience. Um when we do this as scientific research, we are um you know constantly um building up these metrics of um publications and um you know advancing data analytics. But really publications are our metric for scientific progress. But many times when we're uh developing a new scientific advance, especially interventions, how does that translate it to being used in the real world? Uh that happens rarely.

[00:17:38]And so one of the goals of for us for translation at Neat Labs has always been that how can we work with tools that are already scalable that people may already have like their smartphones or like Jason said any wearable that they might already have and uh how can we plug those in instead of having a barrier of highcost tools uh to be made available to people or new fancy um you know sensors that we must develop first before we um you know get this out there. So, so we're very excited more about the uh translation and scalability of the work and um and I think this work has also progressed because duction you've led neat labs as the co-director really looking at how um interventions in psychiatry that um don't have the best um uh remission rates how they the standard of care can be uh improved over time and Um we'd love to hear more of that from you because you really motivated um this kind of work from the team.

[00:18:51]>> Yeah, absolutely. Well, um before I get to that, maybe we we buried the punchline, but how much better did people get after they went through like tell tell us some of the amazing results you found?

[00:19:05]Yeah. So, um we did see that uh 55% of people did achieve remission which is much better than um most uh um trials that are showing 30% remission rates from other standard interventions. Um this comes with the caveat that this is still a study that's not a randomized control study. That is going to be our next step. But I think um 55% remission just with a lifestyle intervention without any um pharmarmacology or invasive work being done is really great and something that people can do from their own home environment, never having to come in to see um a professional just interacting with a coach. Um I think that's um a really a great step to move forward with. Um Jason also um you did some great analyses to um really show that um um that the changes in mood were associated with changes in um improvement in quality of life which is hard to do. also improvements in uh cognition. So um how people were paying attention, their memory, um all of that those metrics improved. And then um a really uh nicely done analysis where Jason showed that um what the lifestyle domain that was being targeted for a given individual when that improved uh then mood also improved uh directly proportional to that and the lifestyle domains that were not being targeted did not improve. Um similarly in that um it's not that if you target sleep then everything else automatically improves and that's why we're seeing the effects that we're seeing. It's that if we target sleep then sleep improves and then depression improves and then um our study ended at a 6 week time point um in terms of the detailed tracking and who knows maybe the other lifestyle factors also improve after that but really it's this targeting and personalization to the lifestyle domain that matters towards improvement in um the depressed state. Um but yeah, we're mostly very excited that uh many people um felt a lot better and achieved healthy status at the end of the study.

[00:21:38]>> Yeah. I mean, I think what I what I think I love about this approach is it can be added on and suitable at multiple stages of depression. You know, I I think of depression across two axes.

[00:21:58]One is how long have you been depressed?

[00:22:01]So, kind of the chronic aspect of it.

[00:22:03]And then how bad is the depression? How severe is it? And often when we think of mild to moderate depression, which is the vast majority of people with depression actually fall into that mild to moderate range, we may not want to recommend medications. you know, we want to recommend things that are safer and and and healthier, but can still be effective. Of course, when we get to the severe depression range, that's often when, you know, we'll say, "Oh, maybe you should be on a medication or really go to more intensive therapy."

[00:22:37]But the mild to moderate range, if we can have a personalized lifestyle intervention that we can deploy that's scalable, that's shown to be effective, that's a that's a great tool. It's an amazing tool that could be offered by primary care doctors who often you know tell a lot of patients come to see their primary care doctor they might not want to see a therapist or a psychologist or a psychiatrist right um so I see this as really valuable at that level and then also in the folks they start a medication they start some other treatment they go into therapy they go from you know severe depression into the mild to moderate range And then what do we do, right? We don't want to add another medication. We don't want to, hey, why don't at this point why don't you get into this lifestyle intervention. Um, so I, you know, I see this as a tool that can be used in multiple areas of behavioral health.

[00:23:38]And, um, we have folks that we treat, we get into remission, and we want to keep them there. We want to say, "Hey, how could you start working on things to help prevent a new depression from coming, a relapse?" And again, you know, you say, "Okay, um, the beauty of this is this can apply to somebody who who doesn't have depression too, right? We're just looking at what are the behavioral and lifestyle predictors of mood fluctuation that actually is relevant to depression, but it's relevant to everybody. We all have mood fluctuations and we can all likely improve and benefit from knowing what are the things that we should do to to help a healthier mood. So that's what I love about this is it's an intervention that can be applied across the spectrum of problems from healthy to uh you know maybe not the most severe depression but almost up to that point.

[00:24:37]Yeah, thank you for um talking about the you know spectrum of depression and um like Dakshin mentioned most of the cases at some point I read the stat that 85% of cases are in the mild to moderate range and um especially if we can use data to empower people to uh get to a better state then also um it prevents um from the condition from worsening further because without treatment sometimes the condition also worsens further.

[00:25:06]And then um we've also talked about uh applications of this project to um non-mood conditions such as chronic pain and uh we talked about that in a previous episode with uh Dr. Herbert.

[00:25:20]And so we um hope to continue to apply this kind of um analytical advance to other types of um chronic conditions as well. and we're excited about that kind of work too. So, um I hope um everyone enjoyed listening to us. Thank you Dr. Nan and Dr. Ramonathan for your expertise and uh you can find us at the uh Neat Labs kneecast which is um available on um YouTube, Spotify and Apple podcast. And so that's our episode six. Thank you everyone.