Predicting Depression among Patients with Diabetes Using Longitudinal Data. A Multilevel Regression Model

Area of Science:

• Healthcare Analytics
• Medical Informatics
• Big Data in Medicine

Background:

• Depression is prevalent and often undiagnosed in diabetes patients, significantly impacting health outcomes, healthcare utilization, and costs.
• Predicting depression in this population is crucial for proactive assessment and intervention, potentially reducing suicide risk.

Purpose of the Study:

• To develop a generalized multilevel regression model for predicting depression severity and major depression in diabetes patients.
• To utilize longitudinal data from a large clinical trial for model development.

Main Methods:

• A 2-level Poisson regression model was developed using selected predictors from 29 candidate factors.
• Depression severity was measured using the Patient Health Questionnaire (PHQ-9) score.
• Model accuracy was evaluated using Root-Mean-Square Error (RMSE) and classification ability for major depression.

Main Results:

• The final model included two time-invariant and 10 time-varying predictors.
• Subject-specific predictions, incorporating historical patient data, demonstrated higher accuracy (RMSE ~4, ROC AUC ~0.9) than population-average predictions.
• Updating the model with newly obtained patient records potentially enhances predictive accuracy and classification.

Conclusions:

• A generalized multilevel regression model effectively predicts depression in diabetes patients.
• Longitudinal patient data and subject-specific predictions yield high predictive ability for depression assessment.