Importance-aware personalized learning for early risk prediction using static and dynamic health data.

Area of Science:

• Medical Informatics
• Artificial Intelligence in Healthcare
• Clinical Decision Support

Background:

• Accurate early risk prediction is crucial for effective medical treatment and healthcare quality.
• Existing methods often rely on long-term dynamic data, neglecting valuable static patient information.
• There is a need for methods that integrate both static and dynamic health data for adaptive early risk prediction.

Purpose of the Study:

• To develop a novel deep learning approach for accurate early clinical risk prediction.
• To adaptively integrate personalized static and dynamic health data.
• To improve the interpretability and accuracy of risk prediction models.

Main Methods:

• Developed an End-to-end Importance-Aware Personalized Deep Learning Approach (eiPDLA).
• Utilized a long short-term memory with temporal attention for sequential dynamic data.
• Incorporated a residual network with correlation attention for static data relationships.
• Employed a multi-residual multi-scale network with an importance-aware mechanism for adaptive feature fusion.

Main Results:

• The eiPDLA significantly outperformed existing methods in early risk prediction (AUC of 0.944 at 1 year).
• Experimental results demonstrated the effectiveness on a real-world Peptic Ulcer Bleeding dataset.
• Case studies confirmed the high interpretability of the prediction results.

Conclusions:

• Combining static and dynamic health data, along with their influencing relationships, is vital for accurate risk prediction.
• The importance-aware mechanism effectively identifies and prioritizes critical features.
• Accurate early risk prediction facilitates timely treatment design and improves patient outcomes.