Interpretable subgroup learning-based modeling framework: Study of diabetic kidney disease prediction.

Area of Science:

• Medical Informatics
• Machine Learning in Healthcare
• Nephrology

Background:

• Diabetic kidney disease (DKD) presents complex heterogeneity, challenging traditional machine learning risk prediction.
• Global models overlook individual patient variations, while subgroup learning often lacks interpretability and efficiency.

Purpose of the Study:

• Introduce the Interpretable Subgroup Learning-based Modeling (iSLIM) framework to address limitations in DKD risk prediction.
• Enhance the accuracy and interpretability of machine learning models for complex diseases.

Main Methods:

• iSLIM integrates expert knowledge with recursive partitioning for DKD subgroup identification within a large EHR dataset (11,559 patients).
• Separate predictive models were developed for each identified subgroup to improve accuracy and maintain interpretability.

Main Results:

• Five clinically relevant DKD subgroups were identified, demonstrating superior predictive performance.
• The iSLIM framework achieved an average sensitivity of 0.8074, outperforming global models by 0.1104.
• Post hoc analyses supported subgroup validity and identified potential DKD risk factors.

Conclusions:

• iSLIM significantly enhances predictive performance over traditional global models for DKD.
• The framework provides interpretable, subgroup-specific risk factors, improving understanding of DKD heterogeneity.
• iSLIM has the potential to increase clinical adoption of machine learning for decision-making.