Using digital phenotyping to classify bipolar disorder and unipolar disorder – exploratory findings using machine learning models

The aims were to investigate 1) differences in smartphone-based data on phone usage between bipolar disorder (BD) and unipolar disorder (UD) and 2) by using machine learning models, the sensitivity, specificity, and AUC of the combined smartphone data in classifying BD and UD. Daily smartphone-based self-assessments of mood and same-time passively collected smartphone data on smartphone usage was available for six months. A total of 64 patients with BD and 74 patients with UD were included. Patients with BD during euthymic states compared with UD in euthymic states had a lower number of incoming phone calls/ day (B: -0.70, 95%CI: -1.37; -0.70, p = 0.040). Patients with BD during depressive states had a lower number of incoming and outgoing phone calls/ day as compared with patients with UD in depressive states. In classification by using machine learning models, 1) overall (regardless of the affective state), patients with BD were classified with an AUC of 0.84, which reduced to 0.48 when using a leave-one-patient-out crossvalidation (LOOCV) approach; similarly 2) during a depressive state, patients with BD were classified with an AUC of 0.86, which reduced to 0.42 with LOOCV; 3) during a euthymic state, patients with BD were classified with an AUC of 0.87, which reduced to 0.46 with LOOCV. While digital phenotyping shows promise in differentiating between patients with BD and UD, it highlights the challenge of generalizing to unseen individuals. It should serve as an complement to comprehensive clinical evaluation by clinicians.