Making sleep behaviors interpretable: adapting the two-process model of sleep regulation to longitudinal Fitbit sleep and activity behaviors for health insights

This study proposes a framework that adapts the neurobiological two-process model of sleep regulation to interpret large-scale longitudinal Fitbit data, successfully mapping wearable behaviors to circadian and homeostatic scores that align with known biological factors and demonstrate significant associations with depression diagnosis and severity.

The Gist

The Big Idea: Turning "Sleep Data" into "Sleep Stories"

Imagine you have a smartwatch (like a Fitbit) that tracks your sleep. It gives you a mountain of raw numbers: "You slept 6 hours, 14 minutes. You woke up 3 times. You took 5,000 steps."

While these numbers are cool, they don't tell you why you feel tired or what to do about it. It's like being handed a list of ingredients (flour, eggs, sugar) without a recipe. You know what you have, but you don't know how to bake the cake.

This paper is about writing the recipe. The researchers took thousands of people's Fitbit data and organized it into two main "buckets" based on how our brains actually control sleep. They call these buckets Process C and Process S.

---

The Two "Sleep Engines"

To understand the paper, you need to know the two engines that drive your sleep, based on a famous theory from neuroscience:

1. Process C (The Circadian Clock): Think of this as your body's internal metronome or a sun-synchronized alarm clock. It tells you when to sleep based on the time of day, the sun rising, and the sun setting. It's about rhythm and timing.
   • The Analogy: This is the conductor of an orchestra. If the conductor is off-beat, the musicians (your body) get confused about when to start and stop.
2. Process S (The Homeostatic Pressure): Think of this as a sleep battery or a pressure cooker. Every hour you are awake, this pressure builds up. The longer you stay awake and the more you move, the "sleepier" you get. When you sleep, the pressure releases, and the battery recharges.
   • The Analogy: This is like a sponge. The longer you leave it out, the more dust (sleep pressure) it collects. When you sleep, you wash the dust off.

What Did the Researchers Do?

They took data from over 32,000 people in the "All of Us" research program (a massive national health study) who wore Fitbits. They looked at millions of days of data.

Instead of just looking at "total sleep time," they created two new scores for every single day:

• The C-Score: How well did your sleep match the sun and your internal clock today?
• The S-Score: How well did your body build up and release sleep pressure today?

They used a statistical "magic trick" (called factor analysis) to figure out which Fitbit habits (like bedtime, steps taken, or napping) mattered most for each score.

What Did They Find?

The new scores made perfect sense when tested against real life:

• The Age Test: As people got older, their "Clock Score" (C) changed (older people tend to go to bed earlier), and their "Pressure Score" (S) got worse (older people often have lighter, more fragmented sleep). The scores matched what doctors already know about aging.
• The Shift Worker Test: People who worked night shifts had terrible "Clock Scores." Their internal metronome was completely out of sync with the sun.
• The Nap Test: People who took naps had lower "Pressure Scores" for the rest of the day. This makes sense! If you nap, you release some of that sleep pressure early, so you aren't as tired at night.
• The Season Test: In the winter, when it gets dark early, people's "Clock Scores" got worse because their bedtime didn't align with the sunset as well as in the summer.

The Depression Connection (The Big Insight)

The researchers wanted to see if these scores could explain why people with depression often have sleep problems.

• The Diagnosis Link: They found that Process S (The Pressure) was the biggest red flag for having depression. People with depression often have a "broken sleep battery." They don't build up enough pressure during the day (maybe because they aren't moving enough), so they can't sleep well at night.
• The Severity Link: However, when looking at how bad the depression was, both the Clock (C) and the Pressure (S) were equally important.

The Takeaway: If you want to fix the risk of depression, focus on building up sleep pressure (move more, avoid naps). If you want to fix the severity of the symptoms, you need to fix both your movement and your daily rhythm.

Why Does This Matter?

Before this paper, if a doctor looked at your Fitbit data, they might just say, "You slept 6 hours, that's not enough."

Now, with this new framework, a doctor (or an app) could say:

"Your Clock Score is fine, but your Pressure Score is low. This means you aren't building up enough sleep drive during the day. Try this: Stop napping and take a brisk walk in the morning to build up that pressure so you can sleep better tonight."

Summary

This paper is a bridge between raw data (steps, sleep times) and biological meaning (circadian rhythms, sleep pressure). It turns a pile of numbers into a clear guidebook for how to fix your sleep, which could eventually help treat diseases like depression. It's like finally getting the instruction manual for your body's sleep engine.

Technical Summary

1. Problem Statement

The proliferation of wearable devices (e.g., Fitbit) has enabled the collection of sleep and activity data at an unprecedented scale (tens of thousands of participants over millions of days). However, a significant challenge remains in interpreting this raw data.

• Lack of Biological Context: Raw metrics (e.g., total sleep time, step count) often lack direct biological relevance, making it difficult to link them to specific physiological mechanisms or actionable clinical interventions.
• Scalability vs. Precision: Gold-standard sleep measurement (EEG/Polysomnography) is not scalable for longitudinal studies, while self-reports lack reliability. Wearable data fills this gap but requires a framework to translate behavioral outputs into neurobiological processes.
• Goal: The authors aim to bridge this gap by mapping wearable-derived behaviors to the Two-Process Model of Sleep Regulation, a well-established neurobiological framework, to create interpretable, actionable scores.

2. Methodology

Cohort and Data Source

• Dataset: Data from the All of Us Research Program (Version 8), comprising 32,292 participants with passively collected Fitbit data.
• Scale: Approximately 15.7 million day-level observations.
• Sources: A mix of "Bring Your Own Device" (BYOD) participants and the "WEAR" sub-study (devices provided by the study).
• Quality Control (QC):
  • Typical Sleep Period (TSP): Used an algorithm to consolidate fragmented sleep into main sleep episodes.
  • Exclusions: Removed days with >1,000 minutes of sleep, extreme activity outliers, and participants with <14 days of data.
  • Specific QC:
    • Circadian (Cb): Required ZIP codes (for sunrise/sunset calculation) and 7+ consecutive day pairs.
    • Homeostatic (Sb): Required 85% confirmed wear time during wakefulness.

Feature Engineering: Mapping to the Two-Process Model

The authors mapped specific Fitbit behaviors ($b$) to either the Circadian Process (C) or the Homeostatic Process (S):

• Circadian Behaviors ($C_b$):
  • Phasing: Bedtime and wake time (clock time); alignment with local sunset/sunrise.
  • Consistency: Night-to-night midsleep difference; Social Jetlag (difference between weekday/weekend midsleep).
• Homeostatic Behaviors ($S_b$):
  • Wakefulness: Duration of wakefulness prior to the sleep episode (accounting for naps).
  • Activity: Total daily steps and peak 10-minute step count (intensity).
  • Staging: Length of the first deep sleep segment and total deep sleep before the first REM segment (selected after sensitivity analysis to ensure they aligned with expected neurobiological patterns despite Fitbit's known staging limitations).

Statistical Modeling: Iterative Factor Analysis

To create day-level scores without losing daily variance (unlike averaging over time), the authors employed a robust Exploratory Factor Analysis (EFA) approach:

1. Sampling: Randomly sampled one day per participant.
2. Iteration: Repeated the sampling and EFA (using Promax rotation) over 5,000 iterations to avoid within-subject correlation bias.
3. Weighting: Calculated the median factor loading for each behavior across all iterations to determine the final weights.
4. Scoring:
   • Created weighted scores for Circadian ($C_b$) and Homeostatic ($S_b$) processes.
   • Orientation: Scores were oriented such that higher values indicate worse behavior/disruption, and lower values indicate better regulation.

Validation and Clinical Association

• Depression Analysis: The scores were tested against Major Depressive Disorder (MDD) diagnosis (derived from EHR ICD codes, PHQ-9, and CIDI-SF surveys) and depression severity (PHQ-9 scores).
• Models: Logistic regression for diagnosis (Case/Control) and linear regression for severity, covarying for age, gender, race, and recording length.

3. Key Results

Score Construction

• $C_b$ (Circadian): Best explained by a 3-factor model:
  1. Consistency (Social jetlag, midsleep stability).
  2. Bedtime alignment (Bedtime timing, sunset alignment).
  3. Wake time alignment (Wake time, sunrise alignment).
• $S_b$ (Homeostatic): Best explained by a 2-factor model:
  1. Activity (Steps, intensity) and prior wakefulness.
  2. Sleep Staging (First deep sleep segment, deep sleep before REM).
  • Note: Prior wakefulness was de-emphasized in the final score compared to activity and staging features.

Validation in Real-World Scenarios

The scores demonstrated expected biological relationships:

• Age: $C_b$ scores decreased (improved) with age until middle age, then stabilized. $S_b$ scores increased (worsened) significantly in older adults (>70), reflecting deep sleep reduction.
• Shift Work: Individuals with median bedtimes >6:00 AM (likely shift workers) had significantly higher $C_b$ scores (worse circadian alignment).
• Seasonality: $C_b$ scores were higher in winter (worse alignment) compared to summer.
• Napping: Days with naps showed significantly higher $S_b$ scores (indicating reduced sleep pressure buildup), and later naps correlated with even higher scores.

Association with Depression

• Diagnosis (MDD):
  • $S_b$ was strongly associated with MDD diagnosis (OR = 1.5, $p < 2e-16$). This supports the "S-deficiency hypothesis" (low activity/sleep pressure buildup).
  • $C_b$ was also associated but with a lower effect size (OR = 1.28).
• Severity (PHQ-9):
  • Both $C_b$ and $S_b$ were equally associated with depression severity ($\beta \approx 0.2$ for both).
  • Specific behaviors like bedtime timing and sleep inconsistency showed high effect sizes for severity.

4. Key Contributions

1. Novel Framework: Successfully adapted the theoretical Two-Process Model to large-scale, passive wearable data, creating distinct $C_b$ and $S_b$ scores.
2. Methodological Innovation: Developed an iterative factor scoring method that preserves day-level granularity while deriving stable weights, avoiding the loss of temporal resolution inherent in traditional averaging.
3. Actionable Interpretability: Transformed abstract wearable metrics into biologically grounded scores that can guide specific interventions (e.g., "reduce naps to improve $S_b$" or "adjust light exposure to improve $C_b$").
4. Clinical Insight: Provided large-scale evidence that while homeostatic disruption ($S_b$) is a stronger predictor of developing depression, both circadian and homeostatic disruptions contribute equally to the severity of the condition.

5. Significance and Limitations

• Significance: This work enables researchers and clinicians to move beyond "how much sleep" to "how well is sleep regulated" at a population scale. It opens the door for longitudinal studies of behavioral interventions and personalized medicine based on specific sleep regulatory deficits.
• Limitations:
  • Sleep Staging Accuracy: Fitbit's sleep staging (specifically deep/REM differentiation) is known to be less accurate than EEG. The authors mitigated this by selecting only features that followed expected biological covariance, but this remains a constraint.
  • Generalizability: The cohort was predominantly White and female; results may vary in more diverse populations.
  • Causality: As an observational study, it identifies associations rather than causal mechanisms, though the biological plausibility is strong.

In conclusion, this paper provides a robust, biologically grounded framework for interpreting wearable sleep data, transforming raw behavioral logs into clinically relevant insights regarding circadian and homeostatic health.