A machine learning-based phenotyping framework identified three distinct metabolic phenotypes in hemodialysis patients, with a simplified six-parameter model achieving an AUC of 0.893-0.919.
Cross-Sectional (n=1,207)
A novel machine learning-based phenotyping framework identified three distinct metabolic profiles in hemodialysis patients, providing a foundation for algorithm-driven individualized care.
Effect estimate: AUC 0.893-0.919
Clinical heterogeneity among hemodialysis patients necessitates precision medicine approaches transcending conventional single-parameter management. Through machine learning analysis of 1,207 maintenance hemodialysis patients, we developed a novel two-tier phenotyping framework integrating unsupervised K-means clustering across 22 clinical indicators with supervised classification using six universally available biomarkers. Five mechanistically informed composite indicators were constructed, including the Middle-Small Molecule Clearance Index (β2-microglobulin reduction ratio × Kt/V) and ferritin–hemoglobin ratio, achieving superior discriminatory capacity over traditional approaches. Three distinct metabolic phenotypes emerged with exceptional stability (Adjusted Rand Index = 0.9181): high retention-inflammatory (19.5%) characterized by dialysis inadequacy and functional iron deficiency, optimal clearance (24.3%) demonstrating superior toxin removal, and intermediate-stable (56.0%) maintaining balanced parameters. The simplified six-parameter model achieved clinically acceptable performance (AUC: 0.893–0.919, accuracy >88%) enabling automated EMR integration. This cross-sectional phenotype discovery represents the foundational step toward precision nephrology, establishing classification frameworks essential for subsequent longitudinal validation studies. The methodology facilitates phenotype-guided interventions: intensified dialysis for high retention-inflammatory patients, clearance optimization for optimal clearance patients, and proactive monitoring for intermediate-stable patients, advancing hemodialysis toward algorithm-driven individualized care with potential to optimize clinical outcomes and resource utilization, pending prospective validation of phenotype-outcome associations.
Chen et al. (Thu,) conducted a cross-sectional in maintenance hemodialysis (n=1,207). Machine learning-based phenotyping framework was evaluated on Model performance for phenotype classification (AUC 0.893-0.919). A machine learning-based phenotyping framework identified three distinct metabolic phenotypes in hemodialysis patients, with a simplified six-parameter model achieving an AUC of 0.893-0.919.