Abstract

Predicting the Likelihood of Remission with Antidepressant Medication in Depression: A Practical Patient-level Machine Learning Approach

Fischer A, Fleming S, Hagan K, Holt-Gosselin B, Schatzberg AF, Williams L
61st Annual Meeting of the American College of Neuropsychopharmacology. 2022.

Background: Less than half of patients with major depressive disorder (MDD) experience symptom remission from their first antidepressant medication (ADM) trial. There is limited empirical data to guide ADM selection for an individual patient. As such, clinicians rely on trial-and-error and evidence from what works for the average patient when treating MDD with first-line ADM, which results in symptom persistence and extended morbidity and mortality for those who do not remit. Using state-of-the-art machine-learning methods, we developed a model that identified patient-specific predictors of likelihood remission (and non-remission) with first-line ADM.

Methods: Recursive feature elimination was used to develop a parsimonious (25 feature) gradient-boosted decision-tree model that predicted binary remission (yes/no) from ADM. Features included pre-treatment clinical, demographic, cognitive, and behavioral variables collected from participants (N = 1008) during the baseline visit of the International Study to Predict Optimised Treatment in Depression (iSPOT-D): a randomized, parallel-model, open-label, repeated-measure, longitudinal 8-week trial assessing response to three commonly prescribed first-line ADMs (sertraline, escitalopram, venlafaxine). Remission was defined as a score < 7 on the 17-item Hamilton Depression Rating Scale (HDRS) at a 12-week follow-up. Models were evaluated based on their accuracy and AUROC for predicting remission on a held-out test set using only measurements collected at baseline. Shapley values were used to ascertain which variables most impacted model estimates at the group and individual (patient) levels.

Results: The trained model exhibited performance significantly exceeding random chance on a held-out test set (AUROC = .64, 95% CI = .55 to .71, p < .001; Accuracy = .63, 95% CI = .56 to .70, p = .010). Of participants who were identified as being at high risk of non-remission by our model (i.e., predicted probability of remission < 20%), model accuracy was .71. Features identified by recursive feature elimination to be of greatest predictive value include measures of pre-treatment: depression symptom severity, anxiety symptom severity, impaired cognitive function (verbal memory and information processing speed), impaired identification of facial emotions (fear and disgust), and agitation (facial and motor). This model was compared to models using common linear modeling approaches with the same number of features (25). Elastic net regularization yielded poorer performance on the test set relative to our approach (AUROC = .59, Accuracy = .55), as did use of simpler downstream models, such as canonical logistic regression (AUROC = .54, Accuracy = .55) and logistic regression with elastic net regularization (AUROC = .59, Accuracy = .61).

Conclusions: Using sophisticated machine-learning methods, we developed a model that identified patient-specific predictors of the likelihood of remission from first-line ADM, which could have powerful clinical implications if implemented in outpatient psychiatry or primary care settings. Moreover, findings illustrate the benefit of using sophisticated machine-learning models (gradient-boosted decision trees) over simple logistic regression, as this method improved both model accuracy and interpretability. Results highlight how variables obtained from standard clinical evaluations may be subjected to machine-learning models to assist in developing individually tailored treatment plans, thereby optimizing outcomes and reducing morbidity and mortality for patients with MDD.