Interrelated feature selection from health surveys using domain knowledge graph.

Area of Science:

• Computational biology
• Health informatics
• Machine learning

Background:

• Prior research often analyzes single-disease datasets, limiting the scope of health advice.
• Identifying patterns across multiple chronic illnesses can reveal shared causes and improve lifestyle guidance.
• Current methods struggle to generalize health advice effectively from diverse patient data.

Purpose of the Study:

• To develop a novel knowledge-based qualitative method for feature selection in multi-disease datasets.
• To improve the predictive power of machine learning models for chronic disease likelihood.
• To enhance the effectiveness of health advice by considering a patient's overall health profile.

Main Methods:

• Constructed a novel knowledge-based qualitative method for feature selection and outlier definition.
• Applied Knowledge Graph-based feature selection to five annual chronic disease health surveys.
• Integrated the method with multiple machine learning and deep learning multi-label classifiers.

Main Results:

• Demonstrated improved classification performance when using Knowledge Graph-based feature selection.
• Successfully removed redundant features and redefined outliers in multi-disease datasets.
• Validated the method's effectiveness across various machine learning and deep learning models.

Conclusions:

• The proposed Knowledge Graph-based feature selection method significantly enhances multi-label classification performance for chronic diseases.
• This approach offers an efficient process for clinicians to select relevant health survey data for personalized care.
• The methodology is adaptable for future advancements, including graph neural networks, for comprehensive health analysis.