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

Area of Science:

• Digital Health
• Computational Psychiatry
• Machine Learning in Healthcare

Background:

• Panic symptoms significantly impact patients with mood and anxiety disorders.
• Predicting panic attacks is crucial for timely intervention and personalized treatment.
• Digital phenotypes offer novel data streams for monitoring mental health.

Purpose of the Study:

• To investigate the predictive power of digital phenotypes and machine learning algorithms for impending panic symptoms.
• To differentiate between the day before panic (DBP) and symptom-free days.
• To identify key digital and clinical indicators associated with panic events.

Main Methods:

• A two-year monitoring study of 43 patients with mood and anxiety disorders.
• Data collection via smartphone applications and wearable devices.
• Analysis of 3,969 data points, including 254 DBP events, using RandomForest, GradientBoost, and XGBoost classifiers.

Main Results:

• The XGBoost model achieved a high predictive performance with a ROC-AUC score of 0.905.
• A simplified model with the top 10 variables also demonstrated strong performance (ROC-AUC 0.903).
• Significant predictors included Childhood Trauma Questionnaire scores, increased step counts, and higher anxiety levels.

Conclusions:

• Machine learning algorithms effectively leverage digital phenotypes for panic symptom prediction.
• Findings support the development of proactive, personalized digital therapies.
• Identified real-life indicators can inform interventions to mitigate panic symptom exacerbation.