JMIR Publications

ABSTRACT

Background:

Developing personalized treatment is one way to improve treatment effectiveness for depressive disorder.

Objective:

This study was aimed to develop and validate the use of machine learning-based prediction models to select personalized pharmacological treatment for patients with depressive disorder.

Methods:

This study used Taiwan's National Health Insurance Research Database. Patients with diagnoses of depressive disorder between 2003 and 2012 were included in this study. The study outcome was treatment failure, which was defined as psychiatric hospitalisation, self-harm hospitalisation, emergency visits, or treatment change. Predictors included the patients’ demographic variables, clinical characteristics of depression, and medical and psychiatric comorbid conditions. Prediction models based on Super Learner algorithms were trained for the initial and the next-step treatment separately. The personalised treatment strategy was developed for choosing the drug with the lowest probability of treatment failure for each patient as the model-selected regimen. We emulated clinical trials to estimate the effect of personalised treatment.

Results:

The areas under the curve of the prediction model using Super Learner was 0·627 for the initial treatment and 0·751 for the next-step treatment. Patients treated with model-selected regimens had reduced treatment failure rates, with a 0·84-fold (95% confidence interval [CI] 0·82-0·86) decrease for the initial treatment and a 0·82-fold (95% CI 0·80-0·83) decrease for the next-step treatment.

Conclusions:

Machine learning-based prediction models for depression outcomes had fair prediction accuracies. In emulating clinical trials, we found the model-selected regimen to be associated with a reduced treatment failure rate. Future randomised controlled trials should be conducted to investigate the effectiveness of the use of machine-learning algorithms in clinical practice.