Data Leakage Inflates CKD Prediction Model Performance by 15%

Machine learning's application to early Chronic Kidney Disease (CKD) detection has garnered considerable interest, yet many published studies present inconsistent and potentially misleading results. A critical gap identified is the lack of systematic evaluation regarding methodological flaws, particularly data leakage, limited access to temporal patient data, and inconsistent reporting of clinical indicators. This research addresses these issues through a systematic literature review of nineteen relevant studies focusing on interpretable machine learning for CKD prediction. The study introduces a structured taxonomy and quantitative scoring framework for information leakage to rigorously assess methodological reliability. The analysis uncovered a strong correlation between data leakage and inflated performance metrics. Specifically, studies with high leakage reported an average accuracy of 95.48%, significantly higher than the 80.2% accuracy observed in leakage-free studies, representing an approximate 15.28% increase. Furthermore, a cross-study feature stability analysis revealed that only a small fraction of predictors are consistently reproducible, with over 80% lacking reliability. These findings strongly suggest that many reported performance improvements in CKD prediction models are artifacts of methodological limitations rather than genuine predictive capabilities. The research underscores the urgent need for more rigorous methodology and transparent reporting in this critical healthcare domain.