Comprehensive Summary
Lee et al. developed DeepGAM, an interpretable deep learning model designed to improve the diagnosis of depression using data from the Heart and Soul Study, which included over a thousand patients with stable coronary heart disease. The model combines a generalized additive model (GAM) with a straight-through estimator (STE) to make its decision-making process more transparent and to identify key biological and clinical predictors. Using 99 psychosocial, clinical, and biochemical variables, DeepGAM surpassed traditional models such as SVMs, logistic regression, and neural networks in diagnostic performance. It also retained nearly the same level of accuracy when reduced to only five major features. These included TNF-alpha, osteopontin, plasma NGAL, NT-proBNP, and the microalbumin-to-creatinine ratio. By providing visual representations of how each factor influenced depression outcomes, DeepGAM demonstrated how interpretability and accuracy can coexist in clinical AI applications.
Outcomes and Implications
This study shows how interpretable AI can make depression diagnosis both clear and more practical for medical use. By identifying biological markers such as TNF-alpha and osteopontin, DeepGAM helps doctors see how physical health and psychological symptoms overlap in patients with heart disease. Even when limited to a few inputs, the model showed strong performance. DeepGAM gives light to a faster and more accessible way to screen for depression, giving it promise for its use in clinical settings. This could aid clinicians in identifying high-risk patients early and shaping a treatment plan around each patient's unique profile. Through additional research with larger/more diverse populations, there is an opportunity for models like DeepGAM could be used in routine mental health screening. This could help fill in the gap between complex data and clinical decision-making.