Predicting Outcomes of Cognitive Behavioral Therapy for Depression Using Data Driven Approaches

Cognitive Behavioural Therapy (CBT) is a well established method for treating clinical depression. However, it is only considered successful for roughly half of those receiving it. Transitioning patients who are unlikely to benefit from CBT to alternative treatments, at an early stage, can significantly enhance their quality of life while alleviating the burden on mental health services. This study focuses on predicting the outcome of CBT using Linear Discriminant Analysis (LDA) models. These models are trained using demographic information, as well as data from depression (PHQ-9), anxiety (GAD-7), and functional impairment (WSAS) questionnaires administered during the initial four CBT sessions. The study aims to determine how the accuracy of the prediction models is influenced by different factors, such as the number of therapy sessions considered, the specific questionnaires used, and the analysis of individual symptoms versus the overall questionnaire scores. The data from over 8000 patients receiving high-intensity CBT (step 3 and 4) is analyzed. Prediction models are trained using different combinations of features and validated across multiple theray sites to ensure generalizability. The best performing model achieved a test accuracy of 69.7% (Odds Ratio OR = 5.20) in predicting whether Reliable and Clinically Significant Improvement (RCSI) was achieved. Interestingly, including baseline scores, demographic information, and WSAS/GAD7 data did not significantly improve the accuracy beyond using PHQ9 responses alone. However, incorporating data from successive therapy sessions consistently improved the prediction accuracy, up to a maximum of four sessions (the limit of the analysis). Surprisingly, analyzing individual symptom trajectories did not offer any advantages over analyzing the total scores of the questionnaires. In conclusion, the study demonstrates that monitoring responses to therapy during the initial sessions can be a valuable tool in predicting treatment outcome. Although there were significant correlations between RCSI and symptom level change, the study found no improvement in the model’s prediction accuracy when considering: (1) individual symptom scores instead of the overall questionnaire score, (2) using GAD-7 and WSAS questionnaires in addition to PHQ-9, or (3) including demographic information.