Smartphone-Tracked Digital Markers of Momentary Subjective Stress in College Students: Idiographic Machine Learning Analysis

Abstract

Background: Stress is an important predictor of mental health problems such as burnout and depression. Acute stress is considered adaptive, whereas chronic stress is viewed as detrimental to well-being. To aid in the early detection of chronic stress, machine learning models are increasingly trained to learn the quantitative relation from digital footprints to self-reported stress. Prior studies have investigated general principles in population-wide studies, but the extent to which the findings apply to individuals is understudied.

Objective: We aimed to explore to what extent machine learning models can leverage features of smartphone app use log data to recognize momentary subjective stress in individuals, which of these features are most important for predicting stress and represent potential digital markers of stress, the nature of the relations between these digital markers and stress, and the degree to which these relations differ across people.

Methods: Student participants (N=224) self-reported momentary subjective stress 5 times per day up to 60 days in total (44,381 observations); in parallel, dedicated smartphone software continuously logged their smartphone app use. We extracted features from the log data (eg, time spent on app categories such as messenger apps and proxies for sleep duration and onset) and trained machine learning models to predict momentary subjective stress from these features using 2 approaches: modeling general relations at the group level (nomothetic approach) and modeling relations for each person separately (idiographic approach). To identify potential digital markers of momentary subjective stress, we applied explainable artificial intelligence methodology (ie, Shapley additive explanations). We evaluated model accuracy on a person-to-person basis in out-of-sample observations.

Results: We identified prolonged use of messenger and social network site apps and proxies for sleep duration and onset as the most important features across modeling approaches (nomothetic vs idiographic). The relations of these digital markers with momentary subjective stress differed from person to person, as did model accuracy. Sleep proxies, messenger, and social network use were heterogeneously related to stress (ie, negative in some and positive or zero in others). Model predictions correlated positively and statistically significantly with self-reported stress in most individuals (median person-specific correlation=0.15-0.19 for nomothetic models and median person-specific correlation=0.00-0.09 for idiographic models).

Conclusions: Our findings indicate that smartphone log data can be used for identifying digital markers of stress and also show that the relation between specific digital markers and stress differs from person to person. These findings warrant follow-up studies in other populations (eg, professionals and clinical populations) and pave the way for similar research using physiological measures of stress.

Keywords: digital biomarker; digital phenotype; machine learning; mobile health; mobile phone; personalized models.

Figure 1

Distribution of the time difference between receiving the initial notification and completing the survey (in minutes).

Figure 2

Daily compliance rate over time. To determine daily compliance rate, we computed the percentage of survey responses that were received on a given participation day relative to the total number of surveys to be sent each day based on our original sample size (224×5=1120). If, for instance, all participants completed 4 out of 5 surveys on their first day of participation, daily compliance equals 80%.

Figure 3

Feature importance ranking for each nomothetic (N) and idiographic (I) model, containing features that appeared in the top 10 features of any model. Numeric values represent the ranking of 1 feature for 1 model. Dark (top) cells represent more important features. N-median and I-median represent the median ranking of a feature across N and I models, respectively, where double values indicate a tie between 2 features. Features are ordered by N-median scores. LASSO: least absolute shrinkage and selection operator; RF: random forest; SVR: support vector regression.

Figure 4

Shapley additive explanations (SHAP) beeswarm plot indicating the relative importance of each feature and the relation between feature values and model prediction for the nomothetic random forest model. COVID-19 is coded as 0=before lockdown and 1=during lockdown. Weekday is coded as 0=weekday and 1=weekend.

Figure 5

Stacked bar plots representing the proportion of participants showing a significant positive (blue bar) or negative (red bar) correlation between feature values and Shapley additive explanations (SHAP) values for each idiographic model. A positive correlation indicates that when a given feature has a higher value, then the model predicts that an individual feels more stressed. A negative correlation indicates that when a given feature has a higher value, then the model predicts that an individual feels less stressed. For instance, the idiographic random forest (I-RF) predicts a lower level of stress at later hours of the day in 53.4% (120/224) of individuals. I-LASSO: idiographic least absolute shrinkage and selection operator; I-SVR: idiographic support vector regression.