Personalized Prognostic Prediction of Treatment Outcome for Depressed Patients in a Naturalistic Psychiatric Hospital Setting: A Comparison of Machine Learning Approaches

Objective: Research on predictors of treatment outcome in depression has largely derived from randomized clinical trials involving strict standardization of treatments, stringent patient exclusion criteria, and careful selection and supervision of study clinicians. The extent to which findings from such studies generalize to naturalistic psychiatric settings is unclear. This study sought to predict depression outcomes for patients seeking treatment within an intensive psychiatric hospital setting and while comparing the performance of a range of machine learning approaches. Method: Depressed patients (N = 484; ages 18-72; 89% White) receiving treatment within a psychiatric partial hospital program delivering pharmacotherapy and cognitive behavioral therapy were split into a training sample and holdout sample. First, within the training sample, 51 pretreatment variables were submitted to 13 machine learning algorithms to predict, via cross-validation, posttreatment Patient Health Questionnaire-9 depression scores. Second, the best performing modeling approach (lowest mean squared error; MSE) from the training sample was selected to predict outcome in the holdout sample. Results: The best performing model in the training sample was elastic net regularization (ENR; MSE = 20.49, R-2 = . 28), which had comparable performance in the holdout sample (MSE = 11.26; R-2 = . 38). There were 14 pretreatment variables that predicted outcome. To demonstrate the translation of an ENR model to personalized prediction of treatment outcome, a patient-specific prognosis calculator is presented. Conclusions: Informed by pretreatment patient characteristics, such predictive models could be used to communicate prognosis to clinicians and to guide treatment planning. Identified predictors of poor prognosis may suggest important targets for intervention.