DeepGAM: An interpretable deep neural network using generalized additive model for depression diagnosis: Data from...

Area of Science:

• Artificial Intelligence
• Machine Learning
• Computational Psychiatry

Background:

• Deep neural networks excel in various tasks but often lack interpretability, hindering clinical applications.
• Estimating depression status from personal medical data is gaining traction, yet understanding model decisions remains challenging.
• Existing black-box models make it difficult to discern the individual impact of input features on depression diagnosis.

Purpose of the Study:

• To introduce DeepGAM, a deep learning-based generalized additive model, enhancing interpretability in depression diagnosis.
• To develop a model that can identify the positive and negative impacts of individual components on depression status.
• To improve feature selection and model interpretability in the context of mental health prediction.

Main Methods:

• Proposed DeepGAM, a deep learning generalized additive model incorporating neural network-based additive functions.
• Designed network architecture and objective function for constrained, regularized, and interpretable outputs.
• Utilized a direct-through estimator (STE) for feature selection via gradient descent, enabling performance with fewer features.

Main Results:

• DeepGAM achieved the highest AUC (0.600) and F1-score (0.387), surpassing traditional neural networks and IGANN.
• STE selected five key features that performed comparably to 99 features, outperforming Lasso and Boruta.
• Demonstrated DeepGAM's interpretability and robust performance on public datasets.

Conclusions:

• DeepGAM with STE offers an accurate and interpretable approach for depression prediction.
• The model provides insights into individual feature contributions, addressing the black-box nature of deep learning.
• This method advances machine learning applications in mental health by combining predictive power with explainability.