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

Background: Sleep disturbance is a core feature of bipolar disorder (BD) and often precedes mood recurrence, particularly in youth. Although actigraphy provides objective sleep measurement, it is limited by cost and monitoring duration. Passive smartphone-based mobile sensing offers a scalable alternative, but its validity in youths with BD is unclear. Methods: Analyses included adolescents and young adults (ages 14-25) with BD-I/II from the PROMPT-BD study with at least four days of concurrent actigraphy and mobile sensing. Actigraphy-derived sleep metrics (total sleep time [TST], sleep onset, sleep offset, midsleep, wake after sleep onset [WASO]) were compared with smartphone-derived proxies (total offline time [TOT], onset, offset, midsleep, phone use after sleep onset [PASO]). Agreement was evaluated using root mean squared error (RMSE) and mixed-effects models. Zero-inflated negative binomial models examined associations between WASO and PASO. Sensitivity analyses tested robustness to missing data, smartphone use patterns, sleep window definitions, operating system, presence vs. absence of mood symptoms and anxiety, and weekend effects. Results: Mobile sensing showed strong convergence with actigraphy for sleep timing and duration (standardized β = 0.54-0.75, all p < .0001). RMSEs were <21 minutes for onset, offset, midsleep, and TST, with strongest agreement for midsleep (RMSE = 14.8 minutes). Mobile sensing slightly overestimated sleep duration and estimated earlier timing. PASO underestimated WASO (RMSE = 48.8 minutes), but greater WASO significantly increased the odds of detecting any PASO (OR per 15 minutes = 1.35, p < .0001). Findings were robust across sensitivity analyses. Conclusions: Passive smartphone-derived sleep metrics approximated actigraphy-based estimates of sleep timing and duration in youth with BD. Given the widespread availability of smartphones in this population, this supports their potential as scalable tools for monitoring circadian disruption and informing early intervention.