Community-based early-stage chronic kidney disease screening using explainable machine learning for low-resource...

Area of Science:

• Medical Informatics
• Machine Learning Applications
• Public Health

Background:

• Early detection of chronic kidney disease (CKD) is crucial for preventing end-stage renal disease.
• Existing screening tools often underperform in South Asia due to differing risk profiles and reliance on high-income country data.
• Many current machine learning (ML) models for CKD use pathology tests and hospital data, limiting community-level applicability in resource-limited settings.

Purpose of the Study:

• To develop and evaluate an explainable ML framework for community-based early-stage CKD screening.
• To create optimized predictor subsets using all available variables and variables excluding pathology tests for practical risk assessment.
• To enable accurate and accessible CKD risk prediction in low-resource settings.

Main Methods:

• A community-based CKD dataset from Bangladesh was utilized to develop predictive models.
• Ten feature selection methods identified robust predictor subsets, and twelve ML classifiers were evaluated.
• External validation was performed on datasets from India, UAE, and Bangladesh, with SHAP used for model interpretation.

Main Results:

• The best ML model achieved 90.4% balanced accuracy using all variables and 89.23% using only non-pathology tests.
• The framework outperformed existing screening tools, requiring fewer and more accessible inputs.
• External validation showed strong generalizability with sensitivities from 78% to 98%, and SHAP identified clinically relevant predictors.

Conclusions:

• Accurate, interpretable, and scalable early-stage CKD screening is feasible using only non-pathology-test features.
• This ML framework shows significant potential for effective community-level CKD screening in resource-constrained environments.
• The study highlights the importance of developing context-specific tools for global health equity in CKD management.