Artificial intelligence and MRI for the prediction of treatment outcome in depression

Depression is a common and debilitating mental health disorder. Although many antidepressant treatments exist, the response to treatment varies substantially between individuals. To expedite the search for effective treatment, there is an urgent need for personalized antidepressant treatment response prediction in depression.
MRI scanning allows for the precise measurement of the structural and functional properties of the brain. Several of such properties are associated with a favorable treatment outcome in large study populations. This work set out to translate findings into personalized predictions using machine learning.
Initial research assessed the impact of segmentation methods on radiomic feature extraction, highlighting significant effects on predictive accuracy. Deep learning-based segmentation methods showed the highest performance.
A key study analyzed MRI and clinical data from 296 patients in a randomized trial, achieving a 68% accuracy in predicting sertraline response, surpassing the 50% chance level of current methods. Multimodal MRI models outperformed single-modality models, with perfusion MRI contributing most to predictions. However, cortical morphometry alone failed to predict treatment outcomes, underscoring the complexity of depression. A novel fMRI analysis method, "Turbulence," showed 70% accuracy in distinguishing responders from non-responders.
This research demonstrates the potential of multimodal MRI and machine learning to improve antidepressant selection. While results are promising, further validation in external and prospective cohorts is needed for clinical implementation. The findings emphasize the importance of developing predictive models to personalize depression treatment and enhance patient outcomes.