Personalized Insights Derived from Wearable Device Data and Large Language Models to Improve Well-Being

By analyzing a year of wearable and self-reported data from 3,139 interns, this study reveals that mood drivers are highly individualized rather than universal, leading to the development of "MoodDriver," a large language model system that generates personalized mental health feedback to advance precision digital interventions.

The Gist

Imagine you have a personal health coach who doesn't just give you a generic pamphlet saying "Exercise more and sleep better." Instead, this coach has been watching you for a whole year, knows your unique habits, and tells you exactly what works for you specifically.

That is the core idea behind this research paper. Here is the story of how they did it, explained simply.

The Problem: The "One-Size-Fits-All" Trap

For a long time, doctors and health experts have told us: "If you want to feel less depressed or anxious, you should walk 10,000 steps and sleep 8 hours."

But the researchers realized this is like telling everyone to wear the same size shoe. It doesn't work for everyone. For some people, walking more makes them feel great. For others, it might make them feel exhausted and grumpy. For some, the time they go to bed matters most; for others, it's how fast their heart beats when they rest.

The old advice was too broad. It ignored the fact that everyone's body and brain are unique.

The Experiment: Watching 3,139 Doctors

To prove this, the researchers studied 3,139 medical interns (doctors in their first year of training). These are people under immense stress, working long hours and night shifts.

They gave everyone a smartwatch (like a Fitbit) and asked them to rate their mood every day on a scale of 1 to 10. They did this for a full year.

What they found was surprising:

• No two people were the same. The things that made one intern happy made another intern sad.
• The "Big Drivers" varied wildly.
  • For 34% of the interns, the most important thing for their mood was simply what time they woke up.
  • For 10.6%, it was how many steps they took.
  • For 20.3%, none of the usual health habits (sleep, steps, heart rate) seemed to affect their mood at all. They were the "outliers" who might need help with something else entirely, like social connection or stress management.
• It changes over time. Sometimes, walking helped a person feel better in January, but by June, it didn't matter as much. The "recipe" for happiness changes as the year goes on.

The Solution: Meet "MoodDriver"

Since we can't give a different doctor to every single person to figure out their unique pattern, the researchers built a robot helper called MoodDriver.

Think of MoodDriver as a super-smart translator.

1. It reads the data: It looks at your smartwatch numbers (steps, sleep, heart rate) and your mood diary.
2. It finds your pattern: It does the math to see, "Oh, for this specific person, waking up later is linked to a better mood," or "For this person, more steps actually lower their mood."
3. It uses a "Brain" (LLM): This is where the Large Language Model (like the AI you are talking to right now) comes in. It takes those dry numbers and turns them into a friendly, personalized email.

What does the email look like?
Instead of saying, "You need to sleep more," it might say:

"Hey! I noticed that on days when you wake up after 7 AM, your mood tends to be higher. But on days when you take more than 5,000 steps, your mood actually drops a little. Maybe you're pushing yourself too hard physically? Let's try focusing on getting enough rest instead."

Or for someone where the data shows no clear link:

"I looked at your data, and interestingly, your steps and sleep don't seem to be the main drivers of your mood right now. This suggests we should look at other things, like how much time you spend with friends or how much sunlight you get."

Why This Matters

This study is a big step toward Precision Mental Health.

• Before: We gave everyone the same map, even though they were trying to go to different destinations.
• Now: We are giving everyone a GPS that knows their specific starting point and destination.

By combining the data from our watches with the conversational power of AI, we can finally stop guessing what makes people feel better and start giving them advice that actually fits their unique lives. It's about moving from "generic health tips" to "your personal health story."

Technical Summary

1. Problem Statement

Health behaviors (e.g., sleep, physical activity) are known to influence mental health, but the relationship is highly heterogeneous across individuals. Current lifestyle interventions rely on population-level recommendations (e.g., "exercise more improves mood"), which often fail because they ignore individual variability.

• The Gap: There is limited evidence on which specific behaviors drive mood for specific individuals and in what contexts.
• The Challenge: Translating complex, high-dimensional longitudinal data from wearables into actionable, personalized insights for users remains difficult. Traditional statistical outputs are not easily interpretable by non-experts.

2. Methodology

A. Study Population and Data Collection

• Cohort: 3,139 medical interns from the Intern Health Study (IHS) (2018–2023), a high-risk population for mood disturbances due to long hours and shift work.
• Data Sources:
  • Wearable Data: Continuous passive monitoring via Fitbit devices (Active minutes, Step count, Resting Heart Rate, Bedtime, Wake-up time, Sleep duration).
  • Self-Reported Data: Daily Ecological Momentary Assessment (EMA) of mood via a mobile app (1–10 scale).
• Preprocessing:
  • Harmonization of activity metrics across different Fitbit API versions (2018–2020 vs. 2021–2023).
  • Filtering outliers (e.g., steps >30k, HR >100).
  • Exclusion of irregular sleep episodes (night shifts) for specific sleep feature analyses to ensure consistency.
  • Inclusion Criteria: Participants required valid mood-feature pairs (≥30 for annual analysis; ≥10 per quarter for temporal analysis).

B. Statistical Analysis

• Correlation Analysis: Spearman rank correlations were computed between each wearable feature and daily mood scores for every individual.
  • Categorization: Associations were classified as Significant Positive (Sig-Pos), Significant Negative (Sig-Neg), or Non-Significant (Non-Sig).
• Temporal Dynamics: The internship year was divided into four quarters to analyze the stability of correlations over time using Sankey plots to visualize transitions between correlation categories.
• Effect Size Quantification: Simple linear regression was performed for significant correlations to calculate individual-level slopes (quantifying how much a unit change in a feature affects mood).

C. The MoodDriver System (LLM Integration)

To translate statistical findings into user-friendly feedback, the authors developed MoodDriver, a Large Language Model (LLM) powered system.

• Input: Time-series data, correlation categories, regression slopes, and data sufficiency metrics.
• Architecture:
  1. Statistical Engine: Computes correlations and regression slopes.
  2. Prompt Engineering: Uses structured prompts with GPT-5 to prevent hallucinations. The prompt constrains the LLM to strictly adhere to the statistical outputs (facts) while generating natural language.
  3. Output Generation: Produces personalized email reports addressing three specific questions:
     • How is my mood affected by my patterns? (Summary of drivers).
     • What does my data look like? (Descriptive statistics).
     • How consistent is my data collection? (Data sufficiency feedback).

3. Key Results

A. Extreme Heterogeneity in Mood Drivers

• Individual Variability: The drivers of mood were highly specific to the individual.
  • Wake-up time was the strongest driver for 34.0% of subjects (mostly positive correlation).
  • Step count was the strongest driver for 10.6% of subjects.
  • Resting Heart Rate (RHR) was the primary negative driver for 15.6% of subjects.
• Directional Variance: The same behavior could have opposite effects on different people. For example, higher step counts were positively associated with mood in 26.9% of subjects but negatively associated in 5.8%.
• Null Findings: 20.3% of participants showed no significant correlations between any of the measured wearable features and their mood, suggesting these specific metrics are not their primary drivers.

B. Temporal Stability vs. Strength Fluctuation

• Stability: While the strength of correlations fluctuated over the four quarters, the direction (positive vs. negative) remained largely stable for individuals who showed significant associations.
• Transitions: Most participants (44.8–62.8%) exhibited temporal variation, primarily transitioning between "Significant" and "Non-Significant" states rather than flipping between positive and negative directions.
• Stable Correlations: Only a small minority (0–6.7%) maintained stable significant correlations across all four quarters.

C. Effect Sizes

• There was substantial variability in the magnitude of effects.
• Averaged Estimates: Across subjects, a 1-point mood improvement was associated with:
  • +38 minutes of active time.
  • +6,600 steps.
  • Going to bed 2 hours earlier.
  • Waking up 4 hours later.
  • Extending sleep by 4 hours.

D. MoodDriver Performance

• The system successfully generated tailored emails for diverse user profiles (e.g., "No Drivers," "Multiple Drivers," "Negative Sleep Correlates").
• The structured prompting effectively minimized LLM hallucinations, ensuring the generated advice was mathematically grounded in the user's specific data.

4. Key Contributions

1. Empirical Evidence of Personalization: Demonstrated that population-level health guidelines are insufficient for a significant portion of the population due to high inter-individual heterogeneity in mood-behavior relationships.
2. Methodological Framework: Established a pipeline combining Digital Phenotyping (wearable data) with LLMs to automate the translation of complex statistical models into personalized, interpretable feedback.
3. MoodDriver System: Developed and validated a scalable, LLM-powered feedback mechanism that addresses the "black box" problem of digital health data by providing specific, data-driven insights (e.g., "Your mood drops when you sleep less than 6 hours").
4. Temporal Insight: Revealed that while the direction of a person's behavioral driver is relatively stable, the significance of that driver fluctuates, suggesting interventions need to be adaptive rather than static.

5. Significance and Future Directions

• Precision Digital Mental Health: This work moves the field from "one-size-fits-all" interventions to precision interventions where recommendations are derived from an individual's unique physiological and behavioral signature.
• Scalability: The use of LLMs allows this personalized feedback loop to be scaled to large populations without requiring manual analysis by clinicians.
• Clinical Implications: Identifying the 20.3% of users with "no drivers" is crucial; it suggests that for these individuals, other factors (social, environmental, nutritional) must be targeted, preventing wasted effort on ineffective behavioral changes.
• Limitations & Future Work:
  • Current data is limited to Fitbit devices and specific features (activity/sleep/HR).
  • The study is observational; future work must test if receiving MoodDriver feedback actually improves mood and behavior (interventional trial).
  • Need to validate the framework in non-intern populations and with other wearable platforms (Apple Watch, Garmin).

Conclusion: The paper successfully demonstrates that combining continuous wearable data with Large Language Models can generate actionable, personalized mental health insights, offering a viable path toward adaptive, evidence-based digital health interventions.