The End of Guesswork? Brain Scans Help Choose the Right Antidepressant

Scientists have used artificial intelligence to try to predict which people with depression would respond best to certain antidepressants. Based on brain scans and symptoms in the first few weeks of treatment, they were able to predict with good accuracy who would improve. This could help in the future to choose the right medication more quickly and avoid failed trials.

Treating depression effectively remains a major challenge. Despite the availability of several antidepressants, more than half of people with major depressive disorder do not improve with the first medication they try.

This means that many patients have to go through several trials before finding a treatment that works, which can take time and worsen their suffering.

One promising alternative to make this process more efficient is the use of machine learning, a form of artificial intelligence that analyzes large volumes of data to identify patterns and make predictions.

The idea is to develop models that, based on clinical information and brain scans, can predict which patients will respond best to a given antidepressant.

However, one of the biggest challenges in this area is ensuring that these models work not only within a single study or hospital, but also across different populations.

Many previous studies have tested these predictions within the same group of patients, but few have assessed whether the results hold up when applied to other clinical settings, with different tests and protocols.

To address this challenge, researchers at Harvard Medical School in the US gathered data from two large clinical studies on depression: EMBARC, conducted in the United States, and CANBIND-1, conducted in Canada.

The study included 363 adult patients (225 from EMBARC and 138 from CANBIND-1) with an average age of 36 years, including 235 women (64.7%).

In both, participants with moderate to severe depression were given a common antidepressant, sertraline or escitalopram, and followed with functional magnetic resonance imaging (fMRI) scans, as well as detailed clinical assessments.

The aim was to test whether models trained on data from one study could predict outcomes in the other. For example, whether a model trained on Canadian data could predict who would improve in the American study, and vice versa.

The researchers also compared different time points and types of information. They found that, rather than relying solely on data collected before treatment began, models that used data from the second week of antidepressant use (such as the initial level of improvement in symptoms) performed better in their predictions.

In addition, including information about connectivity between brain areas, such as the anterior cingulate cortex and the prefrontal cortex, further increased the accuracy of the model.

In this figure, researchers show how certain patterns of brain activity can help predict which patients with depression will respond best to a type of antidepressant. Using brain scan data from people treated with escitalopram (an antidepressant) in the Canadian CANBIND-1 study, the scientists trained a computer model. They then tested this model with data from another study (EMBARC), which used another antidepressant, sertraline, as well as a placebo. The figure shows an area of ​​the brain called the dorsal anterior cingulate (dACC), which stood out as one of the key regions involved in predicting response to treatment. This area was chosen based on data from the study itself and also on previous research that has already pointed to its role in depression. In addition, the graphs show how well the model was able to predict response to treatment in people outside the original study.

The results showed that it is possible to predict, with moderate accuracy, which patients will benefit from antidepressant treatment, even when the model is tested in another clinical trial.

The accuracy was better than chance and showed potential for real-world use. These findings suggest that in the future, it may be possible to personalize depression treatment based on brain scans and early clinical symptoms, shortening response times and avoiding ineffective treatments.

The study represents an important step forward in the path to more accurate and individualized psychiatry.

READ MORE:

Generalizability of Treatment Outcome Prediction Across Antidepressant Treatment Trials in Depression

Peter Zhukovsky, Madhukar H. Trivedi, Myrna Weissman, Ramin Parsey, Sidney Kennedy and Diego A. Pizzagalli

JAMA Netw Open. 2025;8(3):e251310. 

doi:10.1001/jamanetworkopen.2025.1310

Abstract:

Although several predictive models for response to antidepressant treatment have emerged on the basis of individual clinical trials, it is unclear whether such models generalize to different clinical and geographical contexts. To assess whether neuroimaging and clinical features predict response to sertraline and escitalopram in patients with major depressive disorder (MDD) across 2 multisite studies using machine learning and to predict change in depression severity in 2 independent studies. This prognostic study included structural and functional resting-state magnetic resonance imaging and clinical and demographic data from the Establishing Moderators and Biosignatures of Antidepressant Response in Clinical Care (EMBARC) randomized clinical trial (RCT), which administered sertraline (in stage 1 and stage 2) and placebo, and the Canadian Biomarker Integration Network in Depression (CANBIND-1) RCT, which administered escitalopram. EMBARC recruited participants with MDD (aged 18-65 years) at 4 academic sites across the US between August 2011 and December 2015. CANBIND-1 recruited participants with MDD from 6 outpatient centers across Canada between August 2013 and December 2016. Data were analyzed from October 2023 to May 2024. Prediction performance for treatment response was assessed using balanced classification accuracy and area under the curve (AUC). In secondary analyses, prediction performance was assessed using observed vs predicted correlations between change in depression severity. In 363 adult patients (225 from EMBARC and 138 from CANBIND-1; mean [SD] age, 36.6 [13.1] years; 235 women [64.7%]), the best-performing models using pretreatment clinical features and functional connectivity of the dorsal anterior cingulate had moderate cross-trial generalizability for antidepressant treatment (trained on CANBIND-1 and tested on EMBARC, AUC = 0.62 for stage 1 and AUC = 0.67 for stage 2; trained on EMBARC stage 1 and tested on CANBIND-1, AUC = 0.66). The addition of neuroimaging features improved the prediction performance of antidepressant response compared with clinical features only. The use of early-treatment (week 2) instead of pretreatment depression severity scores resulted in the best generalization performance, comparable to within-trial performance. Multivariate regressions showed substantial cross-trial generalizability in change in depression severity (predicted vs observed r ranging from 0.31 to 0.39). In this prognostic study of depression outcomes, models predicting response to antidepressants show substantial generalizability across different RCTs of adult MDD.