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

Patients with bipolar disorder (BD) had 0.70 fewer incoming phone calls per day compared to those with unipolar disorder (UD) during euthymic states.

• During depressive states, patients with BD had fewer incoming and outgoing phone calls per day compared to patients with UD.
• Machine learning models classified patients with BD overall with an area under the curve (AUC) of 0.84, which decreased to 0.48 using a leave-one-patient-out approach.
• In depressive states, the AUC for classifying BD patients was 0.86, reducing to 0.42 with the same cross-validation method.
• During euthymic states, the AUC for classifying BD patients was 0.87, which dropped to 0.46 with leave-one-patient-out cross-validation.
• Digital phenotyping may assist in distinguishing BD from UD, but generalization to new individuals remains challenging.

AI simplified