Machine Learning Predicts Antidepressant Response & Disentangles Drug vs. Placebo

Depression is one of the most widespread mental health disorders worldwide, affecting approximately 4% of the global population. It is characterized by a persistent low mood, disruptions in typical sleeping and/or eating habits, a lack of motivation, a loss of interest in daily activities and unhelpful thought patterns. There are now various treatments for depression, including psychotherapy-based interventions and different types of antidepressant medications. Identifying the best treatment strategy, however, is not always easy, and many patients try different medications before they find one that works for them. Researchers at Stanford University, Lehigh University, the University of Texas at Austin and other institutes explored the potential of machine learning techniques, computational models that can identify patterns in data, for predicting the responses of individual patients to two different antidepressants and to a placebo (i.e., a pill that contains no active chemicals). Their paper, published in Nature Mental Health, showed that machine learning models trained on brain imaging data could predict different patients' responses to the examined drugs with good accuracy. "This work grew out of a practical clinical gap in major depressive disorder (MDD)," Dr. Yu Zhang, Assistant Professor at Stanford University and senior author of the paper, told Medical Xpress. "Many patients do not respond to the first treatment they try, and we still lack reliable biomarkers that can help match patients to treatments. Neuroimaging has shown promise, but most prior biomarkers focused on a single modality alone. We were motivated to integrate brain structure and function in a way that is both predictive and interpretable, and to test whether the resulting biomarkers generalize across independent cohorts." Probing what best predicts patients' responses to treatment Dr. Zhang and his colleagues tried to use machine learning to reliably predict how individual patients would respond to two widely prescribed antidepressant drugs, namely sertraline and escitalopram. These are two selective serotonin reuptake inhibitors (SSRIs), pharmaceutical drugs that increase levels of the neurotransmitter serotonin in the brain by blocking its re-absorption. "We also wanted to explicitly separate medication-related effects from placebo-related effects, since placebo response is substantial in depression trials and can confound biomarker interpretation," said Zhang. The researchers developed a machine learning model on brain imaging data collected from patients diagnosed with MDD, the most severe type of depression, who were prescribed sertraline, escitalopram or a placebo pill. This model gradually identified patterns in the structural organization of the brain and connections between different regions that were associated with an improvement of symptoms following treatment. "Instead of fusing modalities in a generic or unsupervised way, the model learns the shared structure and function signals that matter for a particular clinical prediction," explained Zhang. "We also use strong sparsity regularization, so the model selects a relatively small number of informative connections, which improves interpretability and helps reduce overfitting. Because the model is linear end-to-end, we can map predictions back to specific brain circuits and networks." A shift towards personalized depression treatments Remarkably, the researchers found that their model could predict how well different patients responded to sertraline, escitalopram and a placebo intervention with good accuracy. They also identified measurable patterns in brain scans that were associated with the patients' responses to treatments, uncovering some similarities in how the two SSRIs they examined affect the brain. "Our framework also helps disentangle drug-related response from placebo-related response, which is important for understanding mechanisms and for biomarker validity," said Zhang. "Moreover, the model is designed to be interpretable. It identifies key regions and decomposes the overall predictive pattern into a small number of network constellations with distinct cognitive and personality associations." In the future, the results of this study could inform the development of new strategies to determine what medications are best suited for specific individuals or to identify patients that are likely to respond poorly to SSRIs and other common treatment options. These strategies could shorten or even eliminate the trial-and-error process that many people diagnosed with depression go through after they seek professional help. "Methodologically, we now want to extend the framework to incorporate modality-specific information more explicitly and to handle missing modalities more effectively, which would be important for real world clinical deployment," added Zhang. We are also interested in integrating task-based fMRI to potentially improve prediction and provide additional mechanistic insight. Biologically, we want to study how these structure and function covariation signatures evolve over the course of treatment, including longitudinal change linked to sustained remission and relapse." © 2026 Science X Network