Predicting Antidepressant Response with the STAR*D and CAN-BIND-1 Datasets

Objective: Predict response to antidepressant monotherapy using machine learning techniques, using the CO-MED and STAR*D datasets to train and cross-validate analysis of the CAN-BIND dataset. Part of the CAN-BIND project which has received institutional research ethics board approval at the University of British Columbia

Design: A variety of machine learning techniques will be used to predict antidepressant response comparing between and joining together the level 1 arm of STAR*D, escitalopram + placebo arm of CO-MED, and stage 1 of CAN-BIND-1. Different methods for feature selection will be used to utilize different combinations of clinical and demographic data. We will replicate the methods use in a recent paper which represents the current state of the art, Nie et al 2018.

Analysis plan: The various machine learning and feature selection techniques will be compared using standard measure of accuracy including sensitivity, specificity, PPV, NPV, and AUC. We require a project on NDA so that the group who worked on the prior analyses can share relevant data/algorithms with us.

Data access: We request all clinical and demographic data from the subsets of the clinical trials mentioned above.