Hierarchical Label Distribution Learning for Disease Prediction.

Area of Science:

• Medical Informatics
• Machine Learning
• Computational Biology

Background:

• Accurate disease prediction is vital for early intervention, diagnosis, and treatment, impacting healthcare efficiency and costs.
• Current single-class and multi-class learning methods struggle to differentiate primary from secondary diagnoses, hindering effective treatment strategies.

Purpose of the Study:

• To propose a novel Hierarchical Label Distribution Learning (HLDL) model for disease prediction.
• To enhance fine-grained disease diagnosis by incorporating hierarchical disease classifications and relationships.
• To improve the quantification of diagnoses using label distribution, assigning a degree of description.

Main Methods:

• Developed a novel Hierarchical Label Distribution Learning (HLDL) model.
• Utilized label distribution to quantify diagnostic descriptions.
• Incorporated hierarchical disease classification and inter-disease relationships into the model.

Main Results:

• The HLDL model demonstrated superior performance compared to baseline methods.
• Experimental results on real-world datasets showed statistically significant improvements.
• The model achieved more fine-grained disease predictions.

Conclusions:

• The proposed HLDL model offers a significant advancement in disease prediction accuracy.
• Label distribution learning provides a more nuanced approach to describing diagnoses.
• Considering disease hierarchies and relationships is crucial for effective, fine-grained medical diagnosis.