The Problem We Couldn’t Ignore
Depression affects 21% of US adults and costs society more than $380 billion annually. From person-to-person, there is significant heterogeneity in underlying bio-psycho-social triggers of depression, which may converge to outwardly similar symptom profiles. Given this inter-individual variance, the available treatment pathways for depression only work in about 30% of cases. Further, studies have shown that two-thirds of depression cases remain untreated – patients either lack access to adequate care or do not seek out care due to stigma and disinformation. This scenario begs for development of personalized, affordable and accessible treatment approaches.
It is notable that 80-85% of depression cases are in the category of mild to moderate, but not severe depression. In this cohort, behavioral activation and focus on lifestyle, including optimal sleep, exercise, diet and social connection, can promisingly stimulate and maintain well-being. Yet, when it comes to lifestyle advice, one doesn’t even need a doctor - the internet is full of generic wisdom to sleep better, eat better, exercise more, and socialize more! While such advice may be generally effective at the population level, each individual suffering from depression is unique, with unique lifestyle choices, so naturally there is no one-size-fits-all. Someone whose low mood is driven by social isolation won't be helped by a diet overhaul, and so on. Population-level trials of individual lifestyle attributes yield real results but only because they happen to enroll enough people for whom that particular lever matters. For everyone else, the intervention misses or is not as effective. The question is not whether lifestyle factors affect depression. It's which ones, and for whom. This question is what planted the seed for what would become our Personalized Mood Augmentation (PerMA) trial.
The Precursor to the PerMA study
Our lab name says it all, at the Neural Engineering and Translation Labs (or NEATLabs) translation is the operative word. For us, the end goal of our research is not functioning neurotechnology innovation, but the rigorous implementation and testing of such innovation with patients in the real-world.
This idea drove the development of BrainE, our lightweight digital platform for quantifying cognition and mood in everyday life. In our 2021 proof-of-concept study, we leveraged real-world BrainE assessments with simultaneous smartwatch data, and showed that machine learning models built entirely from a single person's own longitudinal data (N-of-1 ML model) could predict day-to-day fluctuations in depressed mood more accurately than population-level models (Shah et al., 2021). The theory was that if a model can accurately predict your mood, it understands which lifestyle factors are linked to it. Using explainable AI’s most popular SHAP tool (Shapley Additive Explanations), we could reveal exactly how the model predicted each individual's depression, which variables mattered most and the directionality of their relationship with mood. In this way, our 2021 published research laid the groundwork for N-of-1 digital phenotyping of individual mood based on lifestyle as well as neuro-cognitive factors.
What we found confirmed something that is widely understood intuitively but rarely quantified: people have very different underlying biopsychosocial factor determining their mood states.
The PerMA trial
This brings us back to our central ethos, ideally evidence-based research translates to real-world benefit. So the next question we posed was whether we could assign personalized lifestyle interventions based on the different N-of-1 digital phenotypes we observed, thereby, testing whether the mood-lifestyle associations were causal or simply correlative. If exercise-related variables dominated someone's SHAP results, we hypothesized that giving them an exercise intervention would lead to mood improvement. If diet ranked highest, we would target mood-relevant diet modification, and so on. That’s exactly what PerMA did – a unique trial where every person underwent a different intervention course determined by their unique N-of-1 ML model.
How the Study Worked
Phase 1 - Digital Phenotyping (2-4 weeks): 50 adults with mild-to-moderate depression participated in our pilot study. In this phase, participants wore a smartwatch and also responded to brief daily ecological momentary assessments (EMAs) about their mood and lifestyle for up to 60 time points (2-4 times per day, for 2-4 weeks). Then the N-of-1 ML model and SHAP analysis were conducted.
Phase 2 - Individualized Mood Augmentation Plan (iMAP, 6 weeks): Based on the digital phenotype established in Phase 1, each participant received an iMAP targeting their most mood-predictive lifestyle domain, either sleep, exercise, diet, or social connection. Participants then met with a trained behavioral health coach once a week for ~20 min for 6 weeks to review progress on their specific iMAP.
10 of the initial 50 adults enrolled in the study declined the iMAP coaching phase. Of the 40 participants who received iMAPs, 5 were assigned to sleep coaching, 13 to exercise, 5 to diet modification, and 17 to social connection oriented positivity training. With 40 different iMAPs being executed, we are now treating the individual, not the population, and thereby, achieving the end goal of personalized lifestyle medicine.
The Results Speak For Themselves
We observed promising results in the PerMA study –
(1) 55% of study completers no longer met depression criteria, assessed via PHQ9 at post intervention. This is especially notable given the intervention was non-invasive and lifestyle-focused with coach guidance, making this a promising scalable approach. Our observed remission rate was nearly 2X of the ~30% remission rate observed in conventional psychiatric drug and device trials, with the caveat that future randomized controlled trials need to confirm this rate.
(2) Improvements in depressive symptoms were long-lasting and significant even at 12 week post-intervention.
(3) We proved the PerMA trial’s mechanistic specificity by demonstrating that the improvement in depressed mood was directly related to positive changes in the iMAP-targeted lifestyle domain but not in other non-intervention domains.
(4) Notably, cognition in fundamental domains of selective attention, working memory and conflict processing , as well as quality of life metrics significantly improved at post intervention across all study completers.
The Interdisciplinarity of the PerMA Study
PerMA was born at the intersection of disciplines that have rarely been integrated in the past: clinical psychiatry, digital innovation and machine learning engineering. That convergence was not incidental, it emerged from the desire to better quantify what the subjective lived experience of depression tells us.
Getting this interdisciplinary collaboration right took foresight, time and patience, and the lessons learned are shaping everything that comes next. The question was never just how to make such a personalized approach work, but also whether we can build it to work at scale, for anyone, anywhere. That is the challenge we turn to next.
The Road Ahead
As promising as PerMA 1.0 has been, we are actively innovating and improving upon this personalized design. Since its publication, we have now fully automated the PerMA pipeline (data collection, model building, SHAP analysis) on our BrainE digital platform so that other researchers can deploy the same framework for their own questions. We have also expanded the BrainE platform to flexibly support integration of various wearable brands. One ongoing project is applying the personalized approach to chronic pain, using a combination of lifestyle EMAs and Oura ring data, and N-of-1 ML to identify individual chronic pain triggers, in the same way we did for depression. We have further fused PerMA’s predictive AI innovation with generative AI based explanations of the technical SHAP summaries. The gen AI feedback integration has proven very helpful to our clinical partners as well as coaches who are not expected to be well-versed in interpretation of ML outcomes. Notably, we deployed a carefully tested gen AI version trained on our N-of-1 ML dataset from all prior studies instead of using any generic version, thereby, reducing concerns related to gen AI flaws; we continue to monitor gen AI performance for each new N-of-1 prediction model, and if needed iteratively improve its design, to ensure model feedback to clinician and patients remains accurate to the data at all times.
Ultimately, the PerMA study is a successful proof-of-concept of precise measurement informing personalized data modeling and intervention. The PerMA method can be expanded to various types of longitudinal phenotype assessments with greater diversity of actionable predictor variables that can then enrich the personalized intervention landscape. On our end, we remain dedicated to precision clinical translational research that is also accessible to the people who may benefi