Improving Depression Severity Prediction from Passive Sensing: Symptom-Profiling Approach.

Area of Science:

• Digital health
• Mental health informatics
• Computational psychiatry

Background:

• Depression diagnosis relies on subjective methods like interviews and surveys.
• Continuous monitoring of depression is challenging with traditional methods.
• Digital phenotyping offers passive, objective data collection for mental health assessment.

Purpose of the Study:

• To develop and evaluate a novel digital phenotyping approach for detecting depression severity.
• To improve the accuracy of predicting depression severity using passive sensor data.
• To introduce and validate a 'symptom profiling' method for depression assessment.

Main Methods:

• Collected passive sensor data and Patient Health Questionnaire (PHQ-9) scores from 381 participants over three months.
• Developed a 'symptom profile vector' representing nine depressive symptoms, their probabilities, and significance.
• Compared machine learning model performance using sensor features versus symptom profile vectors for predicting depression severity (none/mild, moderate, severe).

Main Results:

• The symptom profiling method significantly improved depression severity prediction accuracy.
• F1 scores increased by up to 0.09, with an average improvement of 0.05.
• The highest achieved F1 score for depression severity prediction was 0.86.

Conclusions:

• Symptom profiling enhances the accuracy of digital phenotyping for depression severity detection.
• Passive sensing combined with symptom profiling offers a promising tool for continuous mental health monitoring.
• This approach can aid in distinguishing depression patterns and improving diagnostic precision.