Passive Sensing of Prediction of Moment-To-Moment Depressed Mood among Undergraduates with Clinical Levels of Depression Sample Using Smartphones

Abstract

Prior research has recently shown that passively collected sensor data collected within the contexts of persons daily lives via smartphones and wearable sensors can distinguish those with major depressive disorder (MDD) from controls, predict MDD severity, and predict changes in MDD severity across days and weeks. Nevertheless, very little research has examined predicting depressed mood within a day, which is essential given the large amount of variation occurring within days. The current study utilized passively collected sensor data collected from a smartphone application to future depressed mood from hour-to-hour in an ecological momentary assessment study in a sample reporting clinical levels of depression (N = 31). Using a combination of nomothetic and idiographically-weighted machine learning models, the results suggest that depressed mood can be accurately predicted from hour to hour with an average correlation between out of sample predicted depressed mood levels and observed depressed mood of 0.587, CI [0.552, 0.621]. This suggests that passively collected smartphone data can accurately predict future depressed mood among a sample reporting clinical levels of depression. If replicated in other samples, this modeling framework may allow just-in-time adaptive interventions to treat depression as it changes in the context of daily life.

Monitoring Changes in Depression Severity Using Wearable and Mobile Sensors

“…Correlations between predicted and observed severity of depressive symptoms ranged from moderate to strong (r ranging between 0.46 and 0.7). The correlation between observed and estimated depression in the time-split model including features from the mobile phone (r = 0.7) was the strongest and was higher than the one of a previous model combining features from the fitbit and from smartphones (the best model yielded an r 2 = 0.44 or r = 0.66) (27) and the one of a model aggregating mobile-based and physiological features (r = 0.58) (26). Notably, despite the high magnitude of the correlations MAE ranged between 3.8 and 4.74 which may be too high of an inaccuracy for the model to be scalable.…”

“…Our study was the first to evaluate behavioral and physiological features, collected entirely passively among a sample of carefully characterized adult individuals with MDD. Previous evaluations of models to estimate depression passively have primarily relied on examining correlations between estimated and observed symptoms (18,26,27). However, indices of associations do not allow a granular evaluation of the accuracy of the models and of the magnitude of the difference between estimated and actual values, impacting scalability.…”

“…Our finding of the role of HRV in predicting depressive symptoms is consistent and extends previous reports. Previous models estimating mood states with moderate accuracy have included measures of heart rate (HR) (27,44) and HRV (26). Moreover, Cao et al (44) reported that in a model including HR features recorded by the fitbit, light exposure, and sleep one of the most important features to estimate mood variation among 18 individuals with MDD was HR.…”

“…Those studies conducted with patients with depression have mostly relied on self-report questionnaires to assign diagnoses and/or measure variation in symptoms and train the models. This has resulted in possible reliability problems, since these investigations have primarily focused on detecting presence or absence of depression rather than on assessing its severity, and they have shown overall low accuracy (14)(15)(16)(17)(18)(23)(24)(25)(26)(27). Moreover, despite evidence showing the existence of physiological indices that are markers of depressed mood (28), that can be continuously collected through wearables (29), and which can be combined with smartphone data to assess mood (27), only two studies have combined these data streams in models to monitor changing depressive symptoms (26,27).…”

Disentangling Trait-Like Between-Individual vs. State-Like Within-Individual Effects in Studying the Mechanisms of Change in CBT

“…Given the omnipresence of smartphones, ecological momentary assessment (EMA) has become increasingly popular in psychological research and holds promise for psychotherapy studies investigating relevant state-like within-person vs. trait-like between-person processes. In addition to EMA, the expansion of “passive” (i.e., no user input required) measurement methods also holds promise for examining predictors and processes of change, including sensor data (e.g., activity levels and movement from accelerometer and GPS, proxies of social interaction from call and text meta-data) from smartphones and wearables ( 43 ), as well as other markers based on motion ( 44 , 45 ), acoustic and language style ( 46 – 48 ) and physiology ( 49 ). The extent to which biological variables, such as hormones ( 50 ), neuroimaging ( 51 – 53 ) and inflammatory biomarkers ( 54 ), provide incremental predictive validity above conventional (and less costly and time-consuming) self-report measures is also an important area of research ( 55 ).…”