Detecting bipolar depression from geographic location data

Abstract

Objective: This paper aims to identify periods of depression using geolocation movements recorded from mobile phones in a prospective community study of individuals with bipolar disorder (BD). Methods: Anonymized geographic location recordings from 22 BD participants and 14 healthy controls (HC) were collected over 3 months. Participants reported their depressive symptomatology using a weekly questionnaire (QIDS-SR16 ). Recorded location data were preprocessed by detecting and removing imprecise data points and features were extracted to assess the level and regularity of geographic movements of the participant. A subset of features were selected using a wrapper feature selection method and presented to 1) a linear regression model and a quadratic generalized linear model with a logistic link function for questionnaire score estimation; and 2) a quadratic discriminant analysis classifier for depression detection in BD participants based on their questionnaire responses. Results: HC participants did not report depressive symptoms and their features showed similar distributions to nondepressed BD participants. Questionnaire score estimation using geolocation-derived features from BD participants demonstrated an optimal mean absolute error rate of 3.73, while depression detection demonstrated an optimal (median±IQR) F1 score of 0.857±0.022 using five features (classification accuracy: 0.849 ± 0.016; sensitivity: 0.839 ± 0.014; specificity: 0.872±0.047). Conclusion: These results demonstrate a strong link between geographic movements and depression in bipolar disorder. Significance: To our knowledge, this is the first community study of passively recorded objective markers of depression in bipolar disorder of this scale. The techniques could help individuals monitor their depression and enable healthcare providers to detect those in need of care or treatment.