An Intelligible Risk Stratification Model Based on Pairwise and Size Constrained Kmeans

Area of Science:

• Computational biology
• Data science
• Public health

Background:

• Clinical disease prevention relies on accurate risk stratification.
• Current risk scores may not reflect individual health conditions or intervention needs.
• Determining the optimal number of risk tiers is often unclear.

Purpose of the Study:

• To develop a novel method for risk stratification that integrates supervised information and size constraints.
• To improve the identification of high-risk populations based on similarity measurements.
• To achieve a balanced and feasible stratification solution.

Main Methods:

• The Pairwise and Size Constrained K-means (PSCKmeans) algorithm was proposed.
• It integrates limited supervised information and size constraints for clustering.
• The method was evaluated on the China Health and Nutrition Survey dataset.

Main Results:

• PSCKmeans naturally stratified diabetes risk into four tiers.
• Achieved 73.8% sensitivity, 85.1% specificity, and a 0.95 ratio of minimum to expected on testing data.
• Outperformed eight previous semi-supervised clustering methods.

Conclusions:

• Semi-supervised clustering with multiple constraints enhances domain-relevant partitioning.
• The PSCKmeans model provides a valuable tool for clinical disease risk stratification.
• This method facilitates targeted interventions for individuals with similar health conditions.