A digital phenotyping dataset for impending panic symptoms: a prospective longitudinal study.

Abstract

This study investigated the utilization of digital phenotypes and machine learning algorithms to predict impending panic symptoms in patients with mood and anxiety disorders. A cohort of 43 patients was monitored over a two-year period, with data collected from smartphone applications and wearable devices. This research aimed to differentiate between the day before panic (DBP) and stable days without symptoms. With RandomForest, GradientBoost, and XGBoost classifiers, the study analyzed 3,969 data points, including 254 DBP events. The XGBoost model demonstrated performance with a ROC-AUC score of 0.905, while a simplified model using only the top 10 variables maintained an ROC-AUC of 0.903. Key predictors of panic events included evaluated Childhood Trauma Questionnaire scores, increased step counts, and higher anxiety levels. These findings indicate the potential of machine learning algorithms leveraging digital phenotypes to predict panic symptoms, thereby supporting the development of proactive and personalized digital therapies and providing insights into real-life indicators that may exacerbate panic symptoms in this population.