Monitoring sleep using smartphone data in a population of college students

Insufficient sleep is linked to adverse health outcomes and is particularly detrimental to mental health, increasing depression and anxiety risk. Monitoring sleep is thus essential for mental health care and for assessing therapeutic response, yet traditional approaches like sleep logs are burdensome and suffer from low engagement. While actigraphy offers accuracy, devices are costly and may not be accessible. Smartphones, widely adopted by nearly 90% of the U.S. population compared to about 20% for wearables, provide a more scalable and potentially equitable approach. Despite capable sensors (motion, location, ambient light, screen state), the Sleep Research Society has considered smartphone-based sleep monitoring premature due to limited validation. Prior research often focuses on wearables or non-smartphone actigraphy, and smartphone studies have faced challenges in reproducibility due to proprietary or unavailable apps. Digital phenotyping—moment-by-moment quantification of behavior via personal devices—offers an opportunity to better capture sleep in situ. This study uses the open-source mindLAMP app to evaluate whether smartphone passive sensor data can estimate sleep duration and relate to validated sleep quality metrics, addressing gaps in scalability, feasibility, and validation.

The authors note that wearables have been emphasized in recent research for sleep assessment but face adoption and equity challenges: only about one-fifth of Americans use them, users skew toward wealthier, more educated, and white demographics, and many abandon devices after weeks. Smartphones are nearly ubiquitous and may better reach high-need populations. Prior smartphone sleep apps have focused on self-reports; studies using smartphones for passive monitoring demonstrated feasibility but lacked replicability due to proprietary software or app unavailability. Some prior methods inferred sleep using environmental cues or calendar assumptions, or used only device usage data, potentially missing activity when screens are off. Bayesian models like SensibleSleep used phone events but did not incorporate accelerometer signals. The authors position a data-driven smartphone-based approach, leveraging both screen state and accelerometer, as a means to improve passive sleep estimation and link to validated outcomes like PSQI.

Study design and participants: Data were drawn from a prior protocol (JMIR Protocols) conducted under IRB approval at Beth Israel Deaconess Medical Center. Participants were recruited via social media to use the open-source mindLAMP app for 28 days. They completed daily activities, daily surveys (sleep duration, sleep quality), and weekly surveys (PHQ-9, GAD-7, PSS, UCLA Loneliness, PSQI, DWAI, TAM-related questions). Of 108 entering enrollment, 34 were discontinued due to inactivity; 74 completed the study; 7 were excluded for prior participation, yielding 67 participants for analysis. Passive data and sleep estimation: Smartphones collected accelerometer and device usage (screen/lock state). Accelerometer data (x, y, z in g) were transformed to jerk (first derivative of acceleration). Otsu's method was applied to jerk magnitude distributions to derive a per-user threshold separating high (active) versus low (inactive) states. A participant was considered active/awake when jerk magnitude exceeded threshold or when device use reported an on-event; other periods were considered inactive/asleep. Missing data segments were concatenated: if missingness was between matching states (asleep→missing→asleep), the missing segment was imputed as that state; for mismatches (asleep→missing→awake), the missing period was considered inactive, assuming the resumption of data indicates a transition to activity. Daily sleep duration estimates were computed as time in inactive state, used as a proxy for time in bed/sleep. Surveys: Daily active sleep duration (prior night, hours) and daily sleep quality (0–10 scale, higher indicating lower quality) were collected; PSQI and mental health measures were weekly. Mixed-effects modeling: Using Python 3.8 and statsmodels, a mixed linear model predicted weekly PSQI with participant as a random intercept (to account for baseline differences). Fixed effects included initial PSQI, weekly-aggregated survey scores, and passive sleep duration estimates. Predictive modeling: A simple linear regression model (scikit-learn) predicted weekly PSQI from weekly-aggregated features: average passive sleep duration estimates and averages of daily sleep duration and sleep quality between consecutive PSQI administrations (to align time windows). Model evaluation used leave-one-out cross-validation (LOOCV), reporting mean absolute error (MAE). Data quality and demographics: Participants (n=67) had mean age 20.0 (SD 2.0); 65.7% female; 56.7% white; average of 28.9 (SD 5.3) daily survey responses across the protocol period.

- Data availability and demographics: 67 participants (mean age 20.0±2.0; 65.7% female; 56.7% white). Average of 28.9±5.3 daily responses. - Correlation of passive estimates with self-reports: Mean passive sleep duration correlated with mean daily self-reported sleep duration at r=0.39 (p<0.05) across the study. When restricting to nights with paired passive and active data, correlation increased to r=0.83 (p<0.05). - Associations with PSQI (weekly): Passive sleep duration estimates negatively correlated with PSQI (r=-0.24, p<0.05), indicating longer passively estimated sleep associated with better sleep quality. Active daily sleep quality (higher score = worse) positively correlated with PSQI (r=0.25, p<0.05). Daily self-reported sleep duration did not correlate with PSQI (p=0.41). - Mixed-effects regression predicting PSQI (fixed effects): • Passive sleep duration estimate coefficient = -0.31 (SE 0.14), p=0.03 (significant negative association). • Active sleep duration coefficient = 0.40 (SE 0.23), p=0.09 (not significant). • Active sleep quality coefficient = 0.21 (SE 0.16), p=0.19 (not significant). - Predictive model performance: Linear model using passive and active features predicted weekly PSQI with mean absolute error MAE=0.93 on LOOCV (PSQI range 0–14), i.e., within about 1 point on average.

The study demonstrates that smartphone-based passive sensing can provide practical, scalable estimates of sleep duration that align closely with self-reported measures when data are concurrently available, and more importantly, relate to validated sleep quality outcomes (PSQI). Passive sleep duration estimates showed stronger and directionally appropriate associations with PSQI than self-reported daily sleep duration, suggesting passive metrics may be more reliable indicators of weekly sleep health. Combining passive and active data enabled accurate prediction of PSQI with sub-point average error, indicating clinical relevance for monitoring and guiding interventions. Given challenges with engagement in sleep logs and limitations of wearables (cost, adoption, equity), smartphone-derived metrics offer accessible, automated, and potentially more consistent data streams for digital phenotyping and clinical discussions. The findings support a validation perspective focused on utility for actionable outcomes (e.g., PSQI) rather than absolute sleep staging accuracy.

Smartphone-based, data-driven estimation of sleep duration using accelerometer-derived activity thresholds and device usage correlates strongly with self-reported sleep when measured concurrently and is significantly associated with weekly PSQI. A simple model integrating passive and active data predicts PSQI within about one point, underscoring clinical utility for scalable sleep monitoring and digital phenotyping. Future work should validate against gold standards (actigraphy, polysomnography), replicate in broader and more diverse populations beyond college students, refine handling of missingness and device nonuse, and explore user prompts or minor protocol adjustments to enhance signal quality.

- Assumption that inactivity equals sleep may misclassify periods of stillness (e.g., sedentary wakefulness). - Dependence on consistent smartphone usage; irregular patterns can reduce validity. - Sample limited to technology-enabled college students, potentially limiting generalizability. - Validation was against PSQI rather than actigraphy or polysomnography; no direct physiological benchmark comparison. - Some missing or non-concurrent active/passive data nights; imputation rules may introduce bias. - Single-study dataset; requires replication in other cohorts and settings.