Multifactorial prediction of depression diagnosis and symptom dimensions

Abstract

While depression is a leading cause of disability, prior investigations of depression have been limited by studying correlates in isolation. A data-driven approach was applied to identify out-of-sample predictors of current depression from adults (N = 217) sampled on a continuum of no depression to clinical levels. The current study used elastic net regularized regression and predictors from sociodemographic, self-report, polygenic scores, resting electroencephalography, pupillometry, actigraphy, and cognitive tasks to classify individuals into currently depressed (MDE), psychiatric control (PC), and no current psychopathology (NP) groups, as well as predicting symptom severity and lifetime MDE. Cross-validated models explained 20.6% of the out-of-fold deviance for the classification of MDEs versus PC, 33.2% of the deviance for MDE versus NP, but -0.6% of the deviance between PC and NP. Additionally, predictors accounted for 25.7% of the out-of-fold variance in anhedonia severity, 65.7% of the variance in depression severity, and 12.9% of the deviance in lifetime depression (yes/no). Self-referent processing, anhedonia, and psychosocial functioning emerged as important differentiators of MDE and PC groups. Findings highlight the advantages of using psychiatric control groups to isolate factors specific to depression.

Keywords: Classification; Psychiatric control; Statistical learning.

Figure 1:

Variable Importance Scores for MDE vs PC (1a) and MDE vs NP (1b) Note: No variable importance plots are generated for the NP vs PC model, as the model added uncertainty to prediction instead of reducing it.

Figure 2:

Variable Importance Scores for the Snaith-Hamilton Pleasure Scale (2a) and the Mood and Anxiety Sensitivity Questionnaire, General Distress Subscale (2b)

Figure 3:

Variable Importance Plot for Lifetime History of Depression (3)