Explainable deep learning for disease activity prediction in chronic inflammatory joint diseases

Area of Science:

• Rheumatology
• Artificial Intelligence
• Computational Biology

Background:

• Chronic inflammatory joint diseases (CIJDs) are complex rheumatic diseases influenced by multiple factors.
• Understanding disease progression and predicting outcomes is challenging due to multifactorial influences.
• The Swiss Clinical Quality Management in Rheumatic Diseases (SCQM) Foundation database provides extensive longitudinal patient data.

Purpose of the Study:

• To develop an explainable multi-task learning model (DAS-Net) for predicting disease activity scores in CIJDs.
• To transform longitudinal patient data into comparable representations for improved analysis.
• To identify patients with similar disease progression patterns.

Main Methods:

• A modular deep learning model incorporating feed-forward networks, LSTMs, and attention layers was developed.
• The model was trained on a large dataset of 19,267 CIJD patients from the SCQM registry.
• Explainability was achieved through feature attribution methods (SHAP, attention, feature weighting) and comparison with expert knowledge.

Main Results:

• DAS-Net demonstrated superior performance in predicting future disease activity scores compared to various baseline models.
• Latent patient representations generated by DAS-Net effectively identified patients with similar disease trajectories.
• Feature attribution analysis provided insights into patient characteristics influencing disease progression.

Conclusions:

• DAS-Net offers a powerful and explainable tool for analyzing complex rheumatic diseases.
• The model's ability to predict disease activity and group similar patients can enhance clinical decision-making.
• This approach advances the use of AI in understanding and managing chronic inflammatory joint diseases.