How Well Do Real-World Clinical And Demographic Characteristics Predict Patient Health Questionnaire-9 Scores Among Patients With Treatment-Resistant Depression?

Background: Many real-world databases do not capture the Patient Health Questionnaire-9 (PHQ-9), a self-reported questionnaire used to measure depression severity. We created a model to predict PHQ-9 severity classification among adults with treatment-resistant depression (TRD) using clinical and demographic characteristics in a database containing PHQ-9 data.

Methods: Adults with likely TRD (i.e. >2 antidepressant treatments in an episode) and ≥1 PHQ-9 record on or after index TRD date were identified in DRG’s combined insurance claims and EHR database between 2013-2018. A predictive model utilizing a random forest classifier used over 90 patient-level covariates, some at multiple time points, to classify depression severity category (PHQ-9 score: 0-9=none/minimal/mild, 10-14=moderate, 15-27=moderately severe/severe). Predicted score and depression severity category were compared in a validation dataset to assess model accuracy.

Results: We identified 2,077 adults with TRD, cumulating to 5,356 PHQ-9 scores. Mean(SD) age was 51.6(15.5) years, 79% were female, and mean PHQ-9 score was 10.1(7.2). The model yielded an accuracy of 63%, almost twice as good as the 33% expected by random chance. For each predicted depression severity category, the median actual observed scores (7, 12, and 18) fell within the range for its respective category.

Conclusions: Using a machine learning predictive modeling approach with patient-level data in a claims and EHR database to predict depression severity among adults with TRD yields an accuracy rate significantly greater than chance. This model may be helpful for risk adjustment, identification of patients for novel TRD treatments, and post hoc comparative analyses of patient cohorts.

This poster was presented at the 32nd annual Psych Congress, held Oct. 3-6, 2019, in San Diego, California.