# Towards causal inference-based antidepressant selection with brain and blood biomarkers

## TL;DR

This study uses brain and blood biomarkers to predict which antidepressants will work best for individuals with depression.

## Contribution

The study introduces a causal inference approach combining brain and blood biomarkers for personalized antidepressant selection.

## Key findings

- A model using 15 multi-modal features predicted remission with an AUC of 0.74 for sertraline, 0.71 for bupropion, and 0.73 for placebo.
- Propensity score-matching and counterfactual prediction models showed 82% accuracy for observed outcomes.
- The selected biomarkers are highly sensitive for guiding antidepressant treatment decisions.

## Abstract

This report sought to employ multi-modal integration of pre-treatment brain (electroencephalogram, resting-state functional magnetic resonance imaging) and blood (immune and metabolic) biomarkers to facilitate causal inference-based treatment selection by virtue of establishing predictability of remission to multi-stage antidepressant treatment. Data from two stages of pharmacotherapy in the ‘Establishing Moderators and Biosignatures of Antidepressant Response for Clinical Care for Depression’ (EMBARC) study from participants with both brain and blood biomarkers were included (N = 197). Participants were initially randomized to sertraline or placebo (Stage 1), and depending on clinical response at week-8, their therapy in Stage 2 was either maintained or switched (to sertraline, if a non-responder to placebo, or to bupropion, if a non-responder to sertraline). Three readily accessible clinical features combined with 15 multi-modal features associated with baseline depression severity predicted stage 2 remission with an AUC of 0.74, 0.71, and 0.73 for sertraline, bupropion, and placebo treatment respectively. Propensity score-matching (causal inference) was conducted across Stage 2 treatment arms, and the same features were used to build an unsupervised model to produce the probability of remission to the given Stage 2 treatment (as factual outcome), as well as the alternative treatment not given (as counter factual). While the accuracy of observed outcomes across treatment arms was 82%, the accuracies of predicted counterfactual (unobserved) outcomes warrant future prospective studies. 16 weeks and associated biomarker-based prediction of counterfactuals suggest that the selected markers are highly sensitive features for guiding antidepressant treatment selection.