Prediction of remission of pharmacologically treated psychotic depression: A machine learning approach.

Background: The combination of antidepressant and antipsychotic medication is an effective treatment for major depressive disorder with psychotic features ('psychotic depression'). The present study aims to identify sociodemographic and clinical predictors of remission of psychotic depression treated with combination pharmacotherapy and determine the accuracy of prediction models.

Methods: Two hundred and sixty-nine participants aged 18 to 85 years with psychotic depression were acutely treated with protocolized sertraline plus olanzapine for up to 12 weeks. Three cross-validated machine learning models were implemented to predict remission based on 74 sociodemographic and clinical variables measured at acute baseline. The optimal model for each method was selected by the average fold C-index. Based on the performance of each method, grouped elastic net (cox) regression was chosen to examine the association of each predictor with remission of psychotic depression.

Results: Of the 269 participants, 145 (53.9 %) experienced full remission of the depressive episode and psychotic features. Multivariable models had 65.1 % to 67.4 % accuracy in predicting remission. In the grouped elastic net (cox) regression model, longer duration of index episode, somatic or tactile hallucinations, higher burden of comorbid physical problems, and single or divorced marital status were independent predictors of longer time to remission. A higher number of lifetime depressive episodes and peripheral vascular or cardiovascular disease were predictors of shorter time to remission.

Conclusions: Future research needs to determine whether the addition of biomarkers to clinical and sociodemographic variables can improve model accuracy in predicting remission of pharmacologically-treated psychotic depression.