Using Domain Knowledge and Data-Driven Insights for Patient Similarity Analytics.

Area of Science:

• Biomedical Informatics
• Clinical Decision Support
• Health Data Science

Background:

• Patient similarity analytics is crucial for identifying comparable patient cohorts.
• Electronic Health Records (EHRs) contain rich data for patient characterization.
• Existing similarity measures may not fully leverage clinical context.

Purpose of the Study:

• To introduce D3K, a novel patient similarity measure integrating domain knowledge and data-driven insights.
• To evaluate D3K's performance against baseline approaches using a large EHR dataset.
• To assess the clinical applicability of D3K in supporting shared decision-making.

Main Methods:

• Constructed patient feature vectors from EHRs of 169,434 patients with diabetes, hypertension, or dyslipidemia (DHL).
• Incorporated domain knowledge for variable discretization (binning) to align with clinical guidelines.
• Compared D3K performance with traditional percentile-based binning and other baseline methods.

Main Results:

• D3K demonstrated superior performance across all seven sub-cohorts compared to baseline methods.
• Domain knowledge-based binning significantly outperformed percentile-based binning.
• High agreement (κ = 0.746) was observed between D3K and physician assessments.

Conclusions:

• D3K is a robust patient similarity measure outperforming existing approaches.
• Domain knowledge integration enhances the clinical relevance of patient similarity analytics.
• D3K shows potential for facilitating shared decision-making in managing cardiometabolic conditions.