ABSTRACT
Introduction
The efficacy of antidepressants for patients with major depressive disorder (MDD) varies from individual to individual, making the prediction of therapeutic outcomes difficult. Better methods for predicting antidepressant outcomes are needed. However, complex interactions between biological, psychological, and environmental factors affect outcomes, presenting immense computational challenges for prediction. Using machine learning (ML) techniques with pharmacogenomics data provides one pathway toward individualized prediction of therapeutic outcomes of antidepressants.
Areas covered
This report systematically reviews the methods, results, and limitations of individual studies of ML and pharmacogenomics for predicting response and/or remission with antidepressants in patients with MDD. Future directions for research and pragmatic considerations for the clinical implementation of ML-based pharmacogenomic algorithms are also discussed.
Expert opinion
ML methods utilizing pharmacogenomic and clinical data demonstrate promising results for predicting short-term antidepressant response. However, predictions of antidepressant treatment outcomes depend on contextual factors that ML algorithms may not be able to capture. As such, ML-driven prediction is best viewed as a companion to clinical judgment, not its replacement. Successful implementation and adoption of methods predicting antidepressant response warrants provider education about ML and close collaborations between computing scientists, pharmacogenomic experts, health system engineers, laboratory medicine experts, and clinicians.
Article highlights
Pharmacogenomic data have been used to predict short-term clinical responses to treatment with antidepressants in people with depression using a variety of machine learning methods.
The results of most studies show that high and generally comparable levels of predictive performance can be achieved using these methods; however, individual studies vary widely regarding the machine learning methods, pharmacogenomic features, non-pharmacogenomic features, validation methods, and study drugs that were used, making direct comparisons between the reviewed studies difficult to conduct.
Few studies included an independent dataset, separate from the original dataset(s) used for algorithm development, for validation of algorithm performance.
Several factors may limit both the validity and clinical utility of the predictions achieved by machine learning models for the treatment of depression with antidepressants, including hidden biases in the data, unpredictable transformations of the data within the algorithms themselves, and the inability of machine learning algorithms to consider important but ‘unseen’ factors that are not specifically accounted for in the input data.
For the treatment of depression with antidepressants, machine learning-based tools may be best viewed as companions to clinical judgment within a shared decision-making framework, as opposed to being a driver of clinical decisions.
Declaration of interest
WV Bobo has been supported by the National Institutes of Health, the Agency for Healthcare Quality and Research, the National Science Foundation, the Myocarditis Foundation, and the Mayo Foundation for Medical Education and Research; and he has contributed chapters to UpToDate, all of which are unrelated to the present work. AP Athreya has been supported by the National Institutes of Health, National Science Foundation, Blue Gator Foundation, Alzheimer’s Association, and the Mayo Foundation for Medical Education and Research. The authors have no other relevant affiliations or financial involvement with any organization or entity with a financial interest in or financial conflict with the subject matter or materials discussed in the manuscript apart from those disclosed.
Reviewer disclosures
Peer reviewers on this manuscript have no relevant financial or other relationships to disclose.