Analysis of depression trajectory patterns using collaborative learning.

Area of Science:

• Mental Health Research
• Health Informatics
• Data Science in Healthcare

Background:

• Depression is a prevalent mental illness, impacting 1 in 10 American adults and is a national health priority.
• Electronic Health Records (EHR) offer a valuable infrastructure for understanding depression's dynamic nature and patient trajectories.
• Depression is the most common mental illness encountered in primary care settings.

Purpose of the Study:

• To analyze depression trajectory patterns within a treatment population using longitudinal EHR data.
• To compare various predictive models for individual depression trajectories to monitor treatment outcomes.
• To identify subgroup patterns in depression severity over time.

Main Methods:

• Utilized longitudinal Patient Health Questionnaire (PHQ)-9 scores from over 3,000 patients undergoing ongoing treatment.
• Modeled individual depression trajectories using smoothing splines and identified subgroup patterns with K-means clustering and collaborative modeling (CM).
• Compared predictive performance of Individual Growth Model (IGM), Mixed Effect Model (MEM), CM, and Similarity-based CM (SCM) for PHQ-9 scores.

Main Results:

• Identified five distinct depression trajectory patterns: stable high, stable low, fluctuating moderate, increasing, and decreasing.
• Collaborative modeling (CM) and similarity-based CM (SCM) demonstrated superior predictive accuracy.
• The root mean square error (rMSE) for SCM was 3.21, significantly lower than IGM (12.53), MEM (5.91), and CM (5.18).

Conclusions:

• Developed a flexible, trajectory-based framework for depression assessment and prognosis utilizing EHR data.
• Collaborative modeling proved more effective than traditional methods for predicting depression trajectories and patient outcomes.
• The framework effectively models population heterogeneity in depression using electronic health record data.