Targeted Prescription of Psychological Interventions for Depression Using Machine Learning
Kristina Dale | February 14, 2020 | Depression
This week's discussion revolves around an article published in 2020 in the Journal of Consulting and Clinical Psychology titled “Targeted Prescription of Cognitive-Behavioral Therapy Versus Person-Centered Counseling for Depression Using a Machine Learning Approach.” In this article, authors Delgadillo and Duhne aimed to identify groups of individuals who respond differently to cognitive behavioral therapy (CBT) or person-centered counseling for depression (CfD) using a targeted prescription algorithm.
What did they do?
The authors used routine care data from patients (n=1,435) already being treated for depression in a primary care setting, where they had received either CBT or CfD. The authors aimed to examine what patient characteristics should be implemented in the development of an algorithm that would prescribe either CBT or CfD to produce the best treatment outcome.
Why did they do it?
The authors highlight that research has suggested that some sorts of therapies (e.g. CBT, Cfd) work through similar mechanisms, which justifies the consideration of patients’ treatment preferences, and may lead to better overall outcomes. As a subcomponent of this research, it has been argued that different patients may actually respond better to treatment when therapies are matched to specific patient attributes, but these attributes have yet to be identified reliably.
How did they do it?
For this study, the authors used data from patients receiving mental health treatment in a primary care setting. The program that patients were recruited from uses a stepped care model, where patients first use self-guided interventions, followed by high intensity therapy interventions for patients who either do not respond to the self-help therapies or who are deemed to be more extreme cases. CBT and CfD were administered to patients by accredited therapists. The authors used total scores from the Patient Health Questionnaire-9 (PHQ-9; a measure of depression symptoms) to evaluate symptom improvement.
To develop an algorithm sensitive to post-treatment “reliable and clinically significant improvement” (RCSI), the authors used propensity score matching for CBT and CfD cases, and logistic regression to identify prognostic indices for each treatment. Using optimal scaling and elastic net regulation in their statistical models, the researchers developed a personalized advantage index (PAI), which was calculated, as explained by the authors, by taking the difference between the two prognostic indices for each case. They then calculated an overall index, a mean prognostic index (MPI).
What did they find?
The authors were able to find six variables that were significantly relevant in both treatment models; age, employment status, disability, and baseline diagnostic scores. Disability predicted better outcomes in CfD, but poorer outcomes in CBT. Patients living in more deprived areas were also predicted to have better outcomes in CfD, and poorer outcomes in CBT.
Specific to CBT, the authors found that ethnicity was significantly predictive of poorer outcomes, such that individuals from minority groups were more likely to have poorer outcomes. Higher baseline anxiety, longer duration of illness, lower outcome expectancy, and individuals not taking antidepressant medication were all predicted to have better outcomes in CfD.
Overall, the authors were able to identify that a subgroup of patients (~30%) responded differently to different interventions.The odds ratio that was computed for the comparison of treatment demonstrated that patients receiving their optimal treatment type were two times more likely to attain reliable and clinically significant improvement. There were no significant differences found between interventions in treatment outcome.
What does it all mean (our take)?
Ultimately, this study boils down to taking the concept of individualized mental health intervention and making it real. We found this study to be absolutely wonderfully done - a high integrity study with so much value! It clearly demonstrates the value of patient data and how, if done properly, we can use our data to inform our future intervention efforts.
Individualized mental health care is not something we need to continue dreaming about...we now have the science and the technology to begin moving towards a much more individualized system. Our system of care at Mindyra is built to help providers accomplish this goal by maximizing the utility of patient behavioral health data, and we are looking forward to having a serious impact on the world of mental health care as we continue to move forward.