Using high-dimensional pharmacogenomics data to predict effective antidepressant treatment response and symptom remission in major depressive disorder patients

Background: Major depressive disorder (MDD) is a highly prevalent, chronic and disabling condition. Antidepressants are the mainstay of treatment with selective serotonin reuptake inhibitors (SSRIs) recommended as first-line treatment. However, antidepressant response rates are dismal with only 35-45% of patients achieving remission after initial agent. Patients with MDD are often exposed to a series of antidepressants in a trial-and-error process in effort to achieve symptom remission or treatment response. We hypothesize that utilization of patients’ electronic health record (EHR) and machine learning methods can improve MDD treatment outcome prediction. Methods: Clinical and pharmacy data were extracted from the UPMC EHR and utilized to examine MDD electronic phenotyping in addition to characterizing antidepressant treatment outcomes including dose changes, treatment sequences, and combinations. In addition, EHR features associated with predicting MDD treatment outcomes were explored. A reproducible pipeline was constructed to yield reproducible results with other data sources, including the addition of PGx data. Results: SSRIs were the most common initial antidepressant class prescribed for MDD patients, followed by SNRIs and NDRIs. The most common initial antidepressant prescribed for patients were SSRIs: sertraline, citalopram, and escitalopram, respectively. Early depression patients, those responding to initial antidepressants, comprised 39.69% of the analysis cohort, while 60.31% of patients required a medication switch or augmentation. The most commonly prescribed two-drug sequence was citalopram then bupropion. When examining the probabilities of transitioning between antidepressant classes, transition probabilities to SSRIs were the highest. The highest performing machine learning model for predicting treatment response was a random forest using the top 25 clinical features (accuracy: 77.21%, F1-score: 87.07%), while the best model for predicting symptom remission was a generalized linear model using the top 25 features (accuracy: 68.16%, F1-score: 33.33%). Discussion: SSRIs are commonly prescribed to patients with MDD, not only as first-line treatment, but are just as likely to be revisited throughout the treatment course. Future directions include assessing the value-add of PGx data in predicting antidepressant treatment response, validating results using EHR data from other health systems with more diverse patient populations, and implementing the prediction model in clinical practice to inform antidepressant treatment selection