Background: Although research shows that more depressed patients respond to combined antidepressants (ADM) and psychotherapy than either alone, many patients do not respond even to combined treatment. A reliable prediction model for this could help treatment decision-making. We attempted to create such a model using machine learning methods among patients in the US Veterans Health Administration (VHA).
Methods: A 2018-2020 national sample of VHA patients beginning combined depression treatment completed self-report assessments at baseline and 3 months (n = 658). A learning model was developed using baseline self-report, administrative, and geospatial data to predict 3-month treatment response defined by reductions in the Quick Inventory of Depression Symptomatology Self-Report and/or in the Sheehan Disability Scale. The model was developed in a 70 % training sample and tested in the remaining 30 % test sample.
Results: 30.0 % of patients responded to treatment. The prediction model had a test sample AUC-ROC of 0.657. A strong gradient was found in probability of treatment response from 52.7 % in the highest predicted quintile to 14.4 % in the lowest predicted quintile. The most important predictors were episode characteristics (symptoms, comorbidities, history), personality/psychological resilience, recent stressors, and treatment characteristics.
Limitations: Restrictions in sample definition, a low recruitment rate, and reliance on patient self-report rather than clinician assessments to determine treatment response limited the generalizability of results.
Conclusions: A machine learning model could help depressed patients and providers predict likely response to combined ADM-psychotherapy. Parallel information about potential harms and costs of alternative treatments would be needed, though, to inform optimal treatment selection.
Keywords: Antidepressant medication; Clinical decision support; Depression; Machine learning; Treatment response; Veterans Health Administration.
