Individuals with depression who do not respond to serotonergic antidepressants are more impaired in reward processing and have greater internalizing symptoms, yet there is no validated self-report method to determine likelihood of treatment resistance based on these symptom dimensions. This case-control study leverages machine learning techniques to identify differences in self-reported anhedonia and internalizing symptom profiles of antidepressant non-responders from responders and healthy controls as an initial proof-of-concept for relating these indicators to medication responsiveness.
Random forest classifiers were built to model differences in a set of 24 anhedonia predictors such that each predictor was a standardized mean item score of a subscale within a validated self-report instrument for measuring anhedonia and other internalizing symptoms. Three groups of adults age 18+ who either (1) had depression and were resistant to serotonergic medication, (2) non-resistant, or (3) non-depressed participated online (N = 393). Multiple models were built to classify across all 3 groups as well as within only the 2 depression groups. Feature selection was applied to each of the classification models based on a measure of predictor importance to improve interpretability. The predictors retained from feature selection were then explored using factor and cluster analysis at the item level to determine empirically driven data structures.
78.1% of the non-resistant group and 67.7% of the medication resistant group reported taking some form of antidepressant medication at the time of this study. Of these individuals, 52.7% in the resistant group and 53.9% in the non-resistant group were taking an SSRI only. Another 14.3% of the resistant and 20.6% of the non-resistant medicated group were taking an SSRI with augmentation, and a total of 92.0% of the resistant and 96.0% of the non-resistant medicated group were taking some form of serotonergic medication. 24 logistic regression analyses examining the interaction effect of medication with each predictor variable regressed on group were evaluated. After correcting for multiple comparisons, no interaction effects remained significant. Therefore, no predictors were removed from the analysis. Accuracies for classifiers using all 24 predictors ranged from .54 to .71. Feature selected models generally retained 3-6 predictors and generated accuracies of .42 to .70. Models performing significantly above chance were the 3-group classifier and the resistant vs. non-resistant classifier. Sensitivity was greatest when comparing non-resistant vs. resistant groups, reaching .82 after feature selection with 3 predictors. Individuals with resistant depression displayed 3 distinct symptom profiles along internalizing dimensions of anxiety, anhedonia, motivation, and cognitive function.
The range of sensitivity achieved for resistant depression in this study was comparable to several previous naturalistic machine learning studies using a catch-all set of sociodemographic, diagnostic, genetic, and self-report clinical predictors. Results should be replicated in a prospective cohort sample and clinical assessment of response to determine predictive validity. However, this study demonstrates validity for using a limited anhedonia and internalizing self-report instrument for distinguishing between antidepressant resistant and responsive depression profiles.