Using digital phenotyping to classify bipolar disorder and unipolar disorder - exploratory findings using machine...

Area of Science:

• Digital phenotyping
• Machine learning in psychiatry
• Mental health diagnostics

Background:

• Bipolar disorder (BD) and unipolar disorder (UD) are distinct mood disorders.
• Differentiating between BD and UD is crucial for effective treatment.
• Objective digital markers are needed to complement subjective clinical assessments.

Purpose of the Study:

• To investigate differences in smartphone usage patterns between BD and UD patients.
• To evaluate the efficacy of machine learning models using smartphone data for classifying BD and UD.
• To assess the sensitivity, specificity, and AUC of these models.

Main Methods:

• Collected daily self-assessments of mood and passive smartphone usage data over six months.
• Included 64 patients with BD and 74 patients with UD.
• Employed machine learning models for classification and analyzed performance using cross-validation.

Main Results:

• Patients with BD showed fewer incoming calls during euthymic states compared to UD.
• Patients with BD exhibited fewer incoming and outgoing calls during depressive states versus UD.
• Machine learning models achieved high AUCs (0.84-0.87) overall but significantly dropped with leave-one-patient-out cross-validation (0.42-0.48).

Conclusions:

• Smartphone usage data can differentiate between BD and UD, particularly in specific affective states.
• The generalization of machine learning models to unseen individuals remains a significant challenge.
• Digital phenotyping shows potential as an adjunct to clinical evaluation for BD and UD.