BACKGROUND

Generalized anxiety disorder (GAD) is a highly prevalent condition. Monitoring GAD symptoms requires substantial time, effort, and cost. The development of digital phenotypes of GAD may enable new scalable, timely, and inexpensive assessments of GAD symptoms.

METHOD

The current study used passive movement data collected within a large national cohort (N = 264) to assess GAD symptom severity.

RESULTS

Using one week of movement data, machine learning models accurately predicted GAD symptoms across a continuum (r = 0.511) and accurately detected those individuals with elevated GAD symptoms (AUC = 0.892, 70.0% Sensitivity, 95.5% Specificity, Brier Score = 0.092). Those with a risk score at the 90 percentile or above had 21 times the odds of having elevated GAD symptoms compared to those with lower risk scores. The risk score was most strongly associated with irritability, worry controllability, and restlessness (individual rs > 0.5). The risk scores for GAD were also discriminant of major depressive disorder symptom severity (r = 0.190).

LIMITATIONS

The current study examined the detection of GAD symptom severity rather than the prediction of GAD symptom severity across time. Furthermore, the instant sample of data did not include nighttime actigraphy, as participants were not asked to wear the actigraphs at night.

CONCLUSIONS

These results suggest that artificial intelligence can effectively utilize wearable movement data collected in daily life to accurately infer risk of GAD symptoms.