Can smartphones detect mood? Investigating temporal dynamics in passive sensing data from bipolar patients

Bipolar disorders are severe and recurrent mental illnesses that are characterized by distinct mood states ((hypo-)mania, depression and euthymia) which are associated with different behavioral patterns. Whereas hypomanic and manic episodes are marked by elevated or irritable mood and increased activity, depression is associated with loss of interest, sadness or decreased activity. These distinct states are often blurred by intra-and interinvidiual variability of symptoms. Trying to better characterize and understand affective states, temporal dynamics gained more interest. While (emotional) inertia, the tendency of a system to persist or change, has been linked to depressive episodes, research on its role in bipolar disorder, particularly during (hypo)manic states, remains sparse. Moreover, behavioral inertia, assessed via passive sensing using smartphone parameters is equally underexplored. This study investigates whether passively sensed smartphone data can reveal distinct dynamic signatures of behavioral inertia during depressive and (hypo)manic episodes in patients with bipolar disorder.

Data were drawn from the BipoSense study (N = 29), including 10,587 patient days with 29 depressive and 20 (hypo)manic episodes during the study period of 12 months. Using the movisensXS app, we passively collected smartphone data on activity, communication, sleep, and phone use of bipolar patients. We computed two dynamic indices: lag-1 autocorrelation (AR) as a measure of behavioral inertia and 14-day moving averages (Mov.AVG) for mean-level behavioral intensity. Multilevel logit models were used to compare AR and Mov.AVG during the first, second, and ≥ third weeks of affective episodes versus euthymia, adjusting for multiple testing using Bonferroni-Holm correction. Random intercepts accounted for within-subject clustering.

Results showed significantly increased AR in activity (e.g., minutes in vehicle, kilometers total) and phone use (e.g., count display on, minutes display on) during depressive episodes, particularly in later weeks (all p ≤ .001). Communication parameters also exhibited elevated AR in later depressive weeks. In contrast, (hypo)manic episodes revealed a more heterogeneous pattern: while AR increased in communication (incoming missed calls, call duration) and sleep (wake-up time), it decreased in phone use (count display on). Mov.AVG confirmed clinically expected patterns: decreased activity and communication during depression, and increased activity, communication, and phone use during mania. Notably, AR and Mov.AVG explained independent variance in episode states, with low correlations between the two indices (r ≤ .146), indicating they capture distinct aspects of behavior.

This study exceeds previous findings by using passive sensing smartphone parameters to investigate behavioral patterns during bipolar episodes. For behavioral inertia consistent patterns during depressive weeks were found, whereas more inconsistent patterns were found during (hypo)manic weeks. AR demontrated lower performance as measure for temporal dynamics than Mov.AVG. Smartphone-based passive sensing of activity may act as an indicator for depression. These results support the use of digital phenotyping to differentiate mood states and highlight the importance of dynamic measures. Future research should validate these findings in larger cohorts. Additionally, digital phenotyping tools should explore the clinical utility of potential dynamic markers for enhancing early detection of affective shifts in bipolar patients, thereby enabling timely, personalized interventions.

References

‘No references’