Prediction of depression risk based on AI art prompts

Despite the growing prevalence of depression, current diagnostic methods sometimes fail to identify individual cases, particularly in the presence of comorbid conditions. This study explores the potential of AI-assisted generative art as a complementary tool for assessing depression risk. A sample of 230 participants (Mage = 21.54 years) used Midjourney, a text-to-image software, to create images of themselves using natural language prompts. Depression risk was measured using the 9-item Beck Depression Inventory, and prompt sentiment was analyzed with a pretrained RoBERTa model. Multiple significant correlations of moderate strength were identified, with negative sentiment emerging as the strongest predictor of depression scores. A random forest classifier achieved 80% sensitivity (AUC = 0.83) in identifying high-risk individuals, with even stronger metrics observed under a stricter cutoff criterion, though this introduces methodological challenges related to class imbalance. These findings suggest that cognitive and emotional biases associated with depression can influence self-expression in generative art, highlighting its potential for capturing depressive symptoms and advancing machine-learning-based screening algorithms in the future.