DGHME: A Hierarchical Hybrid Expert Multi-Task Learning Model for Disease Grouping in Diabetes Complication...

Area of Science:

• Medical Informatics
• Machine Learning in Healthcare
• Diabetes Research

Background:

• Accurate prediction of diabetes complications is crucial for patient care.
• Current single-task learning models fail to utilize correlations between diabetes complications, limiting predictive performance.

Purpose of the Study:

• To propose a novel multi-task learning model, the disease-grouping hierarchical mixed expert model (DGHME), for improved diabetes complication prediction.
• To integrate clinical-pathological grouping knowledge into a hierarchical architecture to capture shared pathophysiological mechanisms.

Main Methods:

• Developed a DGHME integrating clinical-pathological grouping knowledge into a multi-task learning framework.
• Constructed a hierarchical network with self-attention shared experts, group-internal shared experts, task-private experts, and a global shared expert.
• Employed an adaptive gating mechanism and uncertainty-based loss weighting for refined feature learning.

Main Results:

• The DGHME demonstrated superior performance compared to existing multi-task baseline models on a real-world dataset.
• Ablation studies confirmed the model's effectiveness in identifying high-risk patients.
• The DGHME enabled more accurate prediction of diabetes complications.

Conclusions:

• The proposed DGHME effectively leverages correlations among diabetes complications through a hierarchical multi-task learning approach.
• This model enhances the accuracy of identifying high-risk individuals and predicting diabetic complications.
• DGHME offers a promising advancement in leveraging machine learning for diabetes management.