Smartphone-tracked digital biomarkers of stress: An idiographic machine learning perspective (Preprint)

Background:
Stress is an important causal factor in common mental disorders such as burnout and depression. To aid in the early detection of chronic stress, machine learning models are increasingly trained to learn mathematical mappings from digital footprints to self-reported stress. Earlier work has studied general principles in population-wide studies, but the extent to which findings apply to individuals is understudied.

Objective:
We investigated 1) if features of smartphone usage log data (e.g., Messenger application use frequency) are digital biomarkers that can be used to predict momentary subjective stress, 2) if these biomarkers are positively or negatively related to momentary subjective stress (at the group and individual levels), and 3) how accurate these potential digital biomarkers are at recognizing momentary subjective stress on a person-by-person basis in out-of-sample data.

Methods:
Using a large-scale, intensive longitudinal dataset (N = 224, 44,381 observations), we trained machine learning models to predict momentary subjective stress, utilizing explainable artificial intelligence to identify potential digital biomarkers.

Results:
We identified prolonged use of Messenger and Social Network site applications and sleep proxies as valid digital biomarkers. The relationships of these markers with momentary subjective stress as well as predictive accuracy of models differed from person to person. In the majority of individuals, model predictions correlated positively and significantly with self-reported stress.

Conclusions:
Our findings indicate smartphone log data can be utilized as digital biomarkers of momentary subjective stress, but the relationship differs from person to person.