Agreement between smartphone-based mobile sensing and actigraphy sleep metrics in young people with bipolar disorder

This study demonstrates that passive smartphone-based mobile sensing provides a valid, scalable, and robust alternative to actigraphy for monitoring sleep timing and duration in adolescents and young adults with bipolar disorder.

The Gist

Imagine you are trying to figure out exactly when a friend falls asleep and when they wake up. You have two ways to do this:

1. The "Sleep Watch" (Actigraphy): You give them a special, expensive wristwatch that tracks their arm movements like a tiny, silent detective. It's very accurate, but it's bulky, costs a lot of money, and the battery dies quickly. You can only wear it for a week or two before you have to take it off.
2. The "Phone Detective" (Mobile Sensing): You just look at their smartphone. Since almost everyone carries their phone everywhere, the phone knows when it's been unlocked, when the screen went dark, and when it was picked up again. It's free, it's always there, and it can watch them 24/7 for months.

The Big Question:
For young people with Bipolar Disorder (a condition where mood swings can be intense and sleep is often a major problem), does the "Phone Detective" tell the same story as the "Sleep Watch"?

What the Researchers Did:
The team at the University of Pittsburgh took 23 young people (ages 14–25) with Bipolar Disorder and asked them to wear the expensive watch and use their phones normally for a few weeks. They then compared the data from both sources to see if the phone could act as a reliable substitute for the watch.

The Results: A Surprising Match!
Think of the "Sleep Watch" as the gold standard ruler. The researchers found that the "Phone Detective" was actually a pretty good ruler, too.

• Timing: When the watch said the person went to sleep at 11:00 PM, the phone usually guessed it was around 10:45 PM or 11:15 PM. That's a difference of less than 20 minutes on average.
• Duration: The phone was very good at guessing how long they slept.
• The "Middle of the Night" Glitch: The phone was less accurate at counting how many times a person woke up and stayed awake (like tossing and turning). The watch saw 40 minutes of wakefulness, but the phone might only see 10 minutes. This is because if you wake up but don't touch your phone, the phone thinks you are still asleep.
• The "Phone in Bed" Factor: The study found that because these young people kept their phones right next to them (or even in bed), the phone was a very honest witness. If they woke up and checked their phone, the phone knew it.

Why This Matters (The "So What?"):
Bipolar Disorder is like a stormy sea. Sleep problems are often the first wave that signals a storm (a mood episode) is coming.

• The Old Way: Doctors used to rely on patients remembering their sleep or wearing those expensive watches for a few weeks. This is like trying to predict a hurricane by looking at the sky for only 15 minutes.
• The New Way: If a phone can track sleep 24/7 for months, it's like having a weather station that never sleeps. If the phone notices the person's sleep pattern getting weird (like staying up later or waking up more), doctors could get an early warning signal before a major mood crash happens.

The Bottom Line:
The study shows that for young people with Bipolar Disorder, their smartphones are surprisingly good at tracking their sleep patterns, almost as well as the expensive medical watches.

The Catch:
The study was small (only 23 people), and the phone isn't perfect at counting every single time someone wakes up in the middle of the night without touching the device. But, it's a huge step forward. It suggests that in the future, we might not need to buy special watches to monitor mental health; we might just be able to use the device already in our pockets to catch trouble early and help people stay well.

Technical Summary

1. Problem Statement

Sleep disturbance is a core feature of Bipolar Disorder (BD) and a critical early warning sign for mood recurrence, particularly in youth. While actigraphy (wrist-worn motion sensors) is the gold standard for objective sleep measurement, it faces significant barriers to longitudinal clinical use:

• Cost and Compliance: High cost of devices and the burden of wearing a watch 24/7.
• Duration: Typically limited to short-term monitoring (e.g., 1–2 weeks), making it difficult to track mood-related changes over months or years.
• Scalability: Not feasible for large-scale, continuous monitoring in naturalistic settings.

Mobile sensing (passive data collection via smartphone sensors) offers a scalable, low-burden alternative. However, its validity as a proxy for sleep metrics in youth with BD—a population with high rates of sleep irregularity and circadian disruption—had not been empirically established against actigraphy.

2. Methodology

Study Design & Participants

• Source: Data drawn from the PROMPT-BD study (Predicting Recurrence of Mood in Patients with Bipolar Disorder).
• Population: 23 adolescents and young adults (ages 14–25) diagnosed with BD-I or BD-II who were currently in remission (minimal mood symptoms for ≥8 weeks).
• Exclusion: Participants with schizophrenia, autism, IQ <70, or conditions interfering with data collection were excluded.
• Data Collection Period: Concurrent data collection over a 24-month study period, with analysis focused on participants having at least 4 days of valid concurrent data.

Data Acquisition

• Actigraphy (Reference Standard): Participants wore wrist-worn actigraphy watches continuously for 2 weeks at intake and follow-ups. Metrics included:
  • Total Sleep Time (TST)
  • Sleep Onset, Offset, and Midsleep timing
  • Wake After Sleep Onset (WASO)
• Mobile Sensing (Proxy): Participants installed the AWARE app on personal smartphones (iOS/Android) to passively collect:
  • Screen unlocks, activity classification, and location.
  • Derived Metrics:
    • Total Offline Time (TOT) $\approx$ TST
    • Onset/Offset/Midsleep $\approx$ Sleep timing
    • Phone After Sleep Onset (PASO) $\approx$ WASO
• Bias Correction: Because data was aggregated into 1-hour intervals, a systematic bias correction was applied: sleep onset was adjusted by -30 minutes and offset by +30 minutes to account for the "last unlock" vs. "first offline hour" lag.

Statistical Analysis

• Agreement Metrics: Root Mean Squared Error (RMSE) calculated for timing and duration.
• Convergence: Mixed-effects linear models regressed actigraphy metrics on mobile sensing metrics (accounting for repeated measures within participants).
• Zero-Inflated Modeling: Zero-inflated negative binomial regression used for PASO/WASO due to high zero-inflation (many nights with no phone use after sleep onset).
• Sensitivity Analyses: Tested robustness against missing data, self-reported phone habits, expanded sleep windows (7 PM–1 PM), operating system (iOS vs. Android), weekend/weekday effects, and presence of mood/anxiety symptoms.

3. Key Contributions

• First Validation in BD Youth: This is the first study to validate smartphone-derived sleep metrics against actigraphy specifically in a clinical sample of youth with Bipolar Disorder.
• Scalable Proxy Identification: Demonstrates that passive mobile sensing can approximate core sleep parameters (timing and duration) without requiring additional hardware or active user input.
• Contextual Nuance: Identifies that the validity of these proxies is influenced by clinical state (stronger correlation during active symptoms) and temporal context (stronger on weekends).

4. Key Results

Agreement on Sleep Timing and Duration

• Strong Convergence: Mobile sensing showed strong agreement with actigraphy for sleep onset, offset, midsleep, and total sleep time.
  • Standardized $\beta$: 0.54 – 0.75 (all $p < .0001$).
  • RMSE: All metrics were under 21 minutes.
  • Best Performance: Midsleep showed the highest agreement (RMSE = 14.8 minutes).
• Systematic Bias: Mobile sensing slightly overestimated sleep duration and estimated earlier sleep timing compared to actigraphy.
• Error Patterns:
  • Onset error was highest when mobile sensing indicated very early sleep (e.g., 9 PM).
  • Offset error was highest when mobile sensing indicated very late wake times (e.g., 11 AM).
  • Midsleep error remained consistently low across the range.

Agreement on Wake After Sleep Onset (WASO) vs. PASO

• Detection vs. Quantification: While mobile sensing (PASO) underestimated the absolute minutes of WASO (RMSE = 48.8 minutes), it was highly effective at detecting the presence of wakefulness.
• Odds Ratio: For every 15-minute increase in actigraphy-derived WASO, the odds of detecting any PASO increased by 35% (OR = 1.35, $p < .0001$).
• Limitation: Among nights where PASO was detected, the duration of phone use did not linearly correlate with the duration of WASO.

Sensitivity & Moderation Effects

• Operating System: iOS users showed lower RMSE (better agreement) than Android users, though the Android sample was small ($n=5$).
• Weekend vs. Weekday: Associations were significantly stronger on weekends. The authors hypothesize that weekends offer less structured constraints, leading to phone usage patterns that more closely reflect intrinsic circadian rhythms.
• Clinical State:
  • Correlations for midsleep and offset were stronger in participants with concurrent mood symptoms.
  • Correlations for onset and midsleep were stronger in participants with comorbid Generalized Anxiety Disorder (GAD).
• Sleep Window: Expanding the sleep detection window (7 PM–1 PM) increased error and outliers, confirming the necessity of a constrained window (9 PM–11 AM) to avoid misclassifying daytime phone absence as sleep.

5. Significance and Implications

Clinical Utility

• Early Intervention: Since sleep disruption often precedes mood recurrence, the ability to passively monitor sleep via smartphones offers a scalable tool for early detection of impending manic or depressive episodes in BD youth.
• Low Burden: Unlike actigraphy, mobile sensing requires no extra hardware, ensuring high compliance and enabling continuous, long-term monitoring (months to years).

Future Directions

• Just-in-Time Interventions (JITAI): The data supports the development of algorithms that trigger real-time interventions (e.g., prompts to reduce screen time or regulate sleep schedules) when circadian disruption is detected.
• Refinement: Future work should focus on improving the quantification of WASO (currently a weak point) and utilizing machine learning to create individualized baselines.
• Ethical Considerations: As these tools move to clinical care, issues regarding data privacy, algorithmic transparency, and equitable access must be addressed.

Conclusion
The study concludes that passive smartphone-based mobile sensing is a valid, scalable, and feasible method for approximating sleep timing and duration in youth with Bipolar Disorder. While it does not perfectly capture sleep quality (specifically WASO duration), its ability to track circadian rhythms and detect sleep disturbances in naturalistic settings makes it a promising tool for precision prevention and mood recurrence monitoring.