A machine learning approach to predict treatment efficacy and adverse effects in major depression using CYP2C19 and clinical-environmental predictors

Background Major depressive disorder (MDD) is among the leading causes of disability worldwide and treatment efficacy is variable across patients. Polymorphisms in cytochrome P450 2C19 (CYP2C19) play a role in response and side effects to medications; however, they interact with other factors. We aimed to predict treatment outcome in MDD using a machine learning model combining CYP2C19 activity and nongenetic predictors. Methods A total of 1410 patients with MDD were recruited in a cross-sectional study. We extracted the subgroup treated with psychotropic drugs metabolized by CYP2C19. CYP2C19 metabolic activity was determined by the combination of ∗1, ∗2, ∗3, and ∗17 alleles. We tested if treatment response, treatment-resistant depression, and side effects could be inferred from CYP2C19 activity in combination with clinical-demographic and environmental features. The model used for the analysis was based on a decision tree algorithm using five-fold cross-validation. Results A total of 820 patients were treated with CYP2C19 metabolized drugs. The predictive performance of the model showed at best.70 accuracy for the classification of treatment response (average accuracy = 0.65, error = ±0.047) and an average accuracy of approximately 0.57 across all the tested outcomes. Age, BMI, and baseline depression severity were the main features influencing prediction across all the tested outcomes. CYP2C19 metabolizing status influenced both response and side effects but to a lower extent than the previously indicated features. Conclusion Predictive modeling could contribute to precision psychiatry. However, our study underlines the difficulty in selecting variables with sufficient impact on complex outcomes.