Student or educator? Sign up with your school email and get 6 months of Pro free.
CoChat

The DOI matches its Crossref record; 102 of the 102 references it cites were found in a registry or an index; 81 of the 110 sources its claims cite were checked for support.

Personalized machine learning guided intervention for optimizing lifestyle behaviors in depression: a pilot study

Open access

Abstract

Personalized data-driven interventions for depression are much needed. Here, we leveraged N-of-1 machine learning (ML) to optimally target behavioral lifestyle interventions for depression. 50 individuals with mild-to-moderate depression enrolled in the single-arm, open-label Personalized Mood Augmentation (PerMA) pilot clinical trial (NCT05662254). Participants completed a two-week digital monitoring phase using smartphone-based ecological momentary assessments (EMAs, 4×/day) plus smartwatch tracking of mood and lifestyle factors (sleep/exercise/diet/social connection). Personalized ML models were generated from these data to identify lifestyle factors most predictive of individual mood, and results were translated to individualized mood augmentation plans (iMAPs) implemented by participants for six weeks with once-a-week health coach guidance. Intervention completers (n = 40) showed significant reduction in depression symptoms (primary outcome self-rated PHQ9 −3.5 ± 3.8, Cohen’s d = −0.89, CI [−1.25 −0.53], p < 0.001; clinician-rated HDRS −7.2 ± 6.8, d = −1.03, CI [−1.41 −0.65], p < 1E-6) with benefits sustained up to 12-week follow-up. Co-morbid anxiety was also significantly reduced (GAD7: d = −0.85, CI [−1.2, −0.49], p < 0.001) and quality of life improved (d = 0.68, CI [0.33, 1.02], p < 0.001). Additionally, objective cognitive measures impacted in depression including selective attention (d = 0.51, CI [0.18, 0.84], p < 0.001), interference processing (d = 0.53, CI [0.2, 0.85], p < 0.01) and working memory (d = 0.66, CI [0.31, 0.99], p < 0.001) showed significant enhancement. EMA tracking confirmed that improvement in depressed mood was specifically predicted by improvement in individually targeted lifestyles (β = 0.4 ± 0.09, p < 0.0005). Finally, decision algorithms and a large-language-model (LLM) could match human coach-led iMAP assignment with up to 95% accuracy. The PerMA trial presents a personalized lifestyle intervention approach for depression and merits scale-up and RCT testing to establish clinical efficacy. PERMA was registered with ClinicalTrials.gov under registry number NCT05662254.

Overview

CoChat Summary

## TL;DR

This single-arm pilot study tested a personalized, machine-learning-guided lifestyle intervention for 50 people with mild-to-moderate depression. In 40 intervention completers, depression, anxiety, quality of life, mindfulness, and several cognitive measures improved, but the authors note that a randomized controlled trial is still needed.

## Key findings

  • Among trial completers (n = 40), PHQ9 depression scores declined significantly from pre- to post-intervention by −3.5 ± 3.8 with d = −0.89 and p < 1E-13; 22/40 (55%) reached remission (PHQ9 < 5) post-intervention.
  • Depression improvement was sustained at follow-up: 6 weeks d = −1.06 (n = 28) and 12 weeks d = −0.52 (n = 22).
  • Anxiety also improved in completers: GAD7 changed by −2.5 ± 3 with d = −0.85 and p < 1E-5; improvement persisted at 6 weeks (d = −0.59) but not 12 weeks (d = −0.3, p = 0.16).
  • Secondary outcomes improved post-intervention in completers: HDRS d = −1.03, MCS12 d = 0.68, MAAS d = 0.78, and cognition improved for selective attention (d = 0.51), interference processing (d = 0.53), and working memory (d = 0.66); emotion bias did not improve (p = 0.3).
  • EMA analyses found that improvement in depressed mood was linked to improvement in the individually targeted lifestyle domain: the regression coefficient for the target-domain slope was β = 0.4 ± 0.09, p < 0.0005, while the off-target slope was not significant (p > 0.1).
  • Automated assignment methods matched human coach decisions well: the naïve decision algorithm matched in 87.5% of cases (35/40), the LLM in 92.5%, and the fine-tuned decision algorithm in 95%.

## Methods

  • Design: Single-arm, open-label pilot trial (PerMA; NCT05662254) with two phases: about 2 weeks of digital monitoring followed by 6 weeks of individualized coach-guided intervention.
  • Sample/data: 50 adults from the San Diego community with mild-to-moderate depression (PHQ9 5–17; mean age 43.3 ± 16.1, 12 males); data collection ran from Winter 2022–Fall 2024. Participants completed smartphone EMAs and wore a smartwatch; total digital monitoring aimed for 60 EMA sessions.
  • Analysis: Personalized N-of-1 ML models used SHAP to rank lifestyle predictors of depressed mood and guide intervention domain assignment (sleep, exercise, diet, social connection). Outcomes were tested mainly with Friedman repeated-measures tests, signed-rank tests, robust linear regression, ANOVA/paired tests, and effect sizes reported as Cohen’s d with 95% CIs.

## Limitations

  • The paper explicitly states the primary limitation is the lack of a randomized control group, so it cannot establish a causal intervention effect.
  • The authors also state a modest sample size of 40 completers; this limits statistical power and generalizability.
  • Evident from the text: the study is single-arm and open-label, so results may be influenced by expectancy, regression to the mean, or other non-specific effects; follow-up data were also incomplete (30% and 45% missing at 6 and 12 weeks, respectively).

Model-written summary. Check it against the paper before relying on it.

Integrity signals

Registry

Crossref · 10.1038/s44277-026-00062-3

Cites

102 of 102 references verified0 citation errors · 0 not checkable

Statistics

Nothing to recompute.

Figures

Fig. 1 Personalized Mood Augmentation (PerMA) trial flow. A Participant flow is shown from eligibility assessments to completion. B Study design is shown from left to right. Informed consent and eligibility screening was followed by a 2–4 week digital monitoring phase with both smartphone and smartwatch tracking until 60 EMA sessions were completed followed by personalized machine learning model generation, which was then followed by the second phase of individual mood augmentation plan (iMAP) assignment and weekly guided sessions with a health coach for 6 weeks, and subsequent post-intervention assessments and then follow-up assessments at 6 and 12 weeks after end of intervention.
Fig. 1 Personalized Mood Augmentation (PerMA) trial flow. A Participant flow is shown from eligibility assessments to completion. B Study design is shown from left to right. Informed consent and eligibility screening was followed by a 2–4 week digital monitoring phase with both smartphone and smartwatch tracking until 60 EMA sessions were completed followed by personalized machine learning model generation, which was then followed by the second phase of individual mood augmentation plan (iMAP) assignment and weekly guided sessions with a health coach for 6 weeks, and subsequent post-intervention assessments and then follow-up assessments at 6 and 12 weeks after end of intervention.CC BY · PubMed Central
Fig. 2 Results of personalized ML modeling. A ML Model accuracies split by the 4 assigned iMAP domains calculated as (100 – mean ± sd of best-fit model mean absolute percentage error, MAPE). ML model accuracies were not significantly different between domains as confirmed by one-way ANOVA (p > 0.66). B Distribution of iMAPs assigned to the 40 study completers: sleep (5), exercise (13), diet (5), and social connection (17). C Four example Shapley plots showing top 5 predictive features of depressed mood from four distinct participants that were recommended either sleep, exercise, diet or social connection based iMAPs. Shapley dot plots of all ranked feature predictors show directionality of prediction; each dot represents a single datapoint, red dots indicate larger positive feature values while blue dots indicate larger negative feature values. X axis is the Shapley values with center point at 0. Positive Shapley values increase the model output (higher depression), negative Shapley values decrease the model output (lower depression). Each plots has multiple iMAP domain-specific variables appear high in the rankings, and directionality aligned with the goals of the behavioral intervention (i.e., better sleep, more exercise, lower fats and sugars, and more social connection are linked to lower depression).
Fig. 2 Results of personalized ML modeling. A ML Model accuracies split by the 4 assigned iMAP domains calculated as (100 – mean ± sd of best-fit model mean absolute percentage error, MAPE). ML model accuracies were not significantly different between domains as confirmed by one-way ANOVA (p > 0.66). B Distribution of iMAPs assigned to the 40 study completers: sleep (5), exercise (13), diet (5), and social connection (17). C Four example Shapley plots showing top 5 predictive features of depressed mood from four distinct participants that were recommended either sleep, exercise, diet or social connection based iMAPs. Shapley dot plots of all ranked feature predictors show directionality of prediction; each dot represents a single datapoint, red dots indicate larger positive feature values while blue dots indicate larger negative feature values. X axis is the Shapley values with center point at 0. Positive Shapley values increase the model output (higher depression), negative Shapley values decrease the model output (lower depression). Each plots has multiple iMAP domain-specific variables appear high in the rankings, and directionality aligned with the goals of the behavioral intervention (i.e., better sleep, more exercise, lower fats and sugars, and more social connection are linked to lower depression).CC BY · PubMed Central
Fig. 3 Depression (PHQ9) and anxiety (GAD7) outcomes throughout the study period. Mean ± sem scores are plotted in black for trial completers (n = 40) and as dashed grey lines for non-completers (n = 10). A PHQ9 scores are shown across the study period from screening to post-intervention with follow-up scores shown at right at 6 weeks (n = 28) and 12 weeks (n = 22). B GAD7 scores are shown across the study period from pre to post-intervention with follow-ups at right. GS: guided session with GS1 corresponding to the first coach meeting to discuss the ML model results and iMAP with each participant. Stars indicate fdr-corrected signed rank significance across all time points relative to pre-intervention: *: p < 0.05; **p < 0.01; ***p < 0.001.
Fig. 3 Depression (PHQ9) and anxiety (GAD7) outcomes throughout the study period. Mean ± sem scores are plotted in black for trial completers (n = 40) and as dashed grey lines for non-completers (n = 10). A PHQ9 scores are shown across the study period from screening to post-intervention with follow-up scores shown at right at 6 weeks (n = 28) and 12 weeks (n = 22). B GAD7 scores are shown across the study period from pre to post-intervention with follow-ups at right. GS: guided session with GS1 corresponding to the first coach meeting to discuss the ML model results and iMAP with each participant. Stars indicate fdr-corrected signed rank significance across all time points relative to pre-intervention: *: p < 0.05; **p < 0.01; ***p < 0.001.CC BY · PubMed Central
Fig. 4 Pre- vs. post-intervention results for secondary outcomes of subjective behavior (top row) and objective cognition (bottom row) in trial completers (n = 40) shown as box and swarm plots. Box plots show median with lower and upper quartiles as the box edges, whiskers denote the data range (excluding outliers) and the scatter points show individual participant values. Effect sizes with lower and upper 95% confidence intervals in brackets are shown under each measure label. Stars indicate fdr corrected significance. *: p < 0.05; **p < 0.01; ***p < 0.001.
Fig. 4 Pre- vs. post-intervention results for secondary outcomes of subjective behavior (top row) and objective cognition (bottom row) in trial completers (n = 40) shown as box and swarm plots. Box plots show median with lower and upper quartiles as the box edges, whiskers denote the data range (excluding outliers) and the scatter points show individual participant values. Effect sizes with lower and upper 95% confidence intervals in brackets are shown under each measure label. Stars indicate fdr corrected significance. *: p < 0.05; **p < 0.01; ***p < 0.001.CC BY · PubMed Central
Fig. 5 Factors supporting depressive symptom improvement. A Change in post vs. pre PHQ9 depression symptoms were associated with improvement in quality of life; scatter plot are shown with best fit line. B Baseline, i.e., pre-intervention measures associated with remission included anxiety (GAD7, left) and mindfulness (MAAS, right). Box plots show median with lower and upper quartiles as the box edges, whiskers denote the data range (excluding outliers) and the scatter points show individual participant values. Remitters depicted in blue and non-remitters in red. *p < 0.05 fdr-corrected.
Fig. 5 Factors supporting depressive symptom improvement. A Change in post vs. pre PHQ9 depression symptoms were associated with improvement in quality of life; scatter plot are shown with best fit line. B Baseline, i.e., pre-intervention measures associated with remission included anxiety (GAD7, left) and mindfulness (MAAS, right). Box plots show median with lower and upper quartiles as the box edges, whiskers denote the data range (excluding outliers) and the scatter points show individual participant values. Remitters depicted in blue and non-remitters in red. *p < 0.05 fdr-corrected.CC BY · PubMed Central
Fig. 6 EMA metrics during intervention phase 2. A Average slope, i.e., z-scored change in phase 2 relative to phase 1 in depressed mood (grey), target (red), and off-target (blue) lifestyle domains. *** indicates p < 0.001; n.s.: non-significant. B Evolution of z-scored depressed mood (grey), target (red) and off-target (blue) EMAs across all participants, shaded regions reflect sem. wrt: with respect to C Z-scored change in phase 2 relative to phase 1 EMA metrics binned by guided sessions (GS); mean ± sem bars are shown for depressed mood (black), target (red) and off-target (blue) EMAs.
Fig. 6 EMA metrics during intervention phase 2. A Average slope, i.e., z-scored change in phase 2 relative to phase 1 in depressed mood (grey), target (red), and off-target (blue) lifestyle domains. *** indicates p < 0.001; n.s.: non-significant. B Evolution of z-scored depressed mood (grey), target (red) and off-target (blue) EMAs across all participants, shaded regions reflect sem. wrt: with respect to C Z-scored change in phase 2 relative to phase 1 EMA metrics binned by guided sessions (GS); mean ± sem bars are shown for depressed mood (black), target (red) and off-target (blue) EMAs.CC BY · PubMed Central

Verification

Claims the paper makes

26 Valid · 3 Not valid · 6 Not checked · 68 Own results and interpretations

50 more claims · 9 with their cited sources checked

Create a free account to see all 103 claims and to verify the 1 reference still open.

See all claims — free account
  • Probably not a substantial claim. 19
  • It may state two separate claims as one. 36
  • The quote only partly states this claim. 1

Create a free account to see these claims.

Self-check

Percentages, sums and reported differences re-computed from the paper’s own numbers.

Self-check results appear here once an independent check has confirmed them.

Cited works that were retracted

Works this paper cites that were later retracted, from the retraction notices publishers deposit. Major when the notice came out before this paper was published.

1 of 1

Citations

Cites

1 retracted

Show all 92 references

Cited by

See all 1

Metadata

Authors
Nan, J., Purpura, S., Jaiswal, S., Afshar, H., Maric, V., Manchanda, J. K., Taylor, C. T., Ramanathan, D., & Mishra, J.
Affiliations
University of California San Diego (US)
Corresponding author
  • Nan, Jason
  • Mishra, Jyoti
Venue
NPP—Digital Psychiatry and Neuroscience
Published
2026-05-19
Volume
4
Issue
1
ISSN
2948-1570 (ISSN-L)
OpenAlex
W7161581526
Cited by
1
References
92
Citations by year
2026: 1
FWCI
5.88
Type
journal article
Language
en
Open access
Licence
Locations
  • NPP—Digital Psychiatry and Neuroscience (publishedVersion, cc-by)
  • PubMed Central (publishedVersion, cc-by)
  • PubMed (publishedVersion)
  • DOAJ (DOAJ: Directory of Open Access Journals) (submittedVersion, cc-by-sa)
Funders
Hope for Depression Research Foundation (N/A)
Topics
  • Social Sciences › Psychology › Applied Psychology › Digital Mental Health Interventions
  • Social Sciences › Psychology › Experimental and Cognitive Psychology › Mental Health Research Topics
  • Social Sciences › Psychology › Experimental and Cognitive Psychology › Anxiety, Depression, Psychometrics, Treatment, Cognitive Processes
Keywords
depression, personalized interventions, lifestyle modification, ecological momentary assessment, health coaching, large language models, anxiety comorbidity, quality of life, cognitive function, sleep, physical activity, diet, social connectedness, digital mental health, PHQ-9, Hamilton Depression Rating Scale, GAD-7, working memory, selective attention

About these checks

Registry, retraction, Cites and statistics results are computed, not judged, with one exception: a printed reference that no index lists is read once by a model before it counts as a citation error. Claims are extracted by a model and kept only when their quote appears verbatim in the text; each cited source is then checked, with a model’s help, for whether it says what the claim relies on. Those verdicts can be wrong — read the claim’s quote and the cited source before relying on one.

The claim verdicts this report counts are experimental: their accuracy is not yet measured.

A language model drew the claims from the paper on 2026-10-09, under version 3 of the claim definition (model deepseek-v4.1-flash, prompt extract-v3).

A second model labelled each claim on 2026-10-09 (model jev-1.13-20260917, definition revision gold-guideline-v0.2).

Cited sources are checked with a language model’s help (gpt-6-sol).

Statistics and arithmetic are recomputed, not judged.

Confidence bands are shown only for fields calibrated on our annotated test set. None is calibrated yet, so no band is shown.

What counts as a claim ›How we measure confidence ›How we check this ›

Are you an author, or do you see a problem with this page? Request a review

How to cite

Cite CoChat for the checks on this page, and the paper for its own findings.

This page
CoChat. (2026). Citation check: Personalized machine learning guided intervention for optimizing lifestyle behaviors in depression: a pilot study. https://app.cochat.ai/papers/10.1038/s44277-026-00062-3
The paper
Nan, J., Purpura, S., Jaiswal, S., Afshar, H., Maric, V., Manchanda, J. K., Taylor, C. T., Ramanathan, D., & Mishra, J. (2026). Personalized machine learning guided intervention for optimizing lifestyle behaviors in depression: a pilot study. NPP—Digital Psychiatry and Neuroscience. https://doi.org/10.1038/s44277-026-00062-3

Data sources

Bibliographic data: Semantic Scholar, OpenAlex, Crossref.

Contains information from the Semantic Scholar Academic Graph Datasets, made available under the ODC Attribution License (ODC-BY 1.0).

Semantic Scholar’s data is described in: Kinney, R., et al. (2023). The Semantic Scholar Open Data Platform. arXiv:2301.10140.