Effectiveness of Artificial Intelligence-Based Approaches in Predicting the Progression of Chronic Kidney Disease: A Systematic Review of Randomized Controlled Trials and Primary Studies

Abstract

Introduction: Chronic kidney disease (CKD) is a progressive disorder with heterogeneous trajectories poorly forecasted by conventional regression-based instruments. Artificial intelligence (AI), encompassing machine learning, tree-based ensembles, deep neural networks, and dynamic Bayesian architectures, has been proposed to exploit high-dimensional, non-linear, and longitudinal clinical data for individualised prognostication. This systematic review evaluates AI-based approaches for predicting CKD progression across the full translational pathway from model derivation to randomised implementation.

Methods: Systematic review conducted per PRISMA 2020 guidelines. Eligible designs were RCTs, cluster RCTs, cohort, case-control, cross-sectional, and observational studies evaluating AI/ML models for predicting CKD progression, kidney failure, or renal-function decline. Risk of bias assessed with PROBAST, RoB 2, and Newcastle-Ottawa Scale. Thirteen pre-specified outcome domains were analysed using structured narrative synthesis.

Results: Thirty-five primary studies comprising >2.2 million individuals, >490,000 images, and >89,000 patient-years met inclusion criteria: 7 randomised trials, 26 cohort/case-cohort studies, and 2 post-hoc analyses. Median AUC for CKD progression was 0.87 (range 0.74-0.99). AI significantly outperformed conventional reference instruments: XGBoost AUC 0.892 versus KFRE 0.745; Klinrisk outperformed KDIGO heatmap at every interval (P<0.01); KDpredict super-learner achieved 5-year IPA 27.8% versus 18.1% for KFRE in Denmark (30.5% versus 14.2% in Scotland). Deep learning on biopsy images achieved AUCs 0.878-0.904 versus 0.786-0.811 for pathologist-estimated scoring. Discrimination was preserved on cross-ethnic external validation (AUC 0.867 and 0.871 in ELSA/HRS). However, randomised implementation evidence was discordant: decision support significantly improved SBP (-14.6 versus -11.7 mmHg, P=0.005), guideline-concordant prescribing (49.9% versus 34.6%, P<0.001), and annualised eGFR loss (0.01 versus 0.95 mL/min/1.73m²/year, P<0.001), whereas passive AI risk display produced no measurable benefit (39% versus 40%, P=1.00). Calibration deteriorated in advanced CKD, with 5-year KFRE overpredicting by 10-18%.

Discussion: AI-based models consistently improve CKD progression discrimination beyond conventional regression, with gains greatest where data are high-dimensional, non-linear, longitudinal, or image-based. Explainability analyses converged on a physiologically coherent predictor set (eGFR, albuminuria, age, creatinine, albumin, haemoglobin, potassium, frailty). Nevertheless, calibration drift over longer horizons, incomplete PROBAST reporting, and sparse randomised evidence that predictive accuracy translates into patient benefit temper enthusiasm. Workflow integration, actionable recommendations, and clinician engagement, rather than algorithmic accuracy alone, determine clinical effectiveness.

Conclusion: AI-based approaches provide significant and clinically meaningful improvements in CKD progression prediction relative to conventional equations, with preservation on external validation. Translation into improved patient outcomes, however, is contingent on active integration of risk estimates into concrete clinical action pathways. Nephrology services should prioritise externally validated, well-calibrated, interpretable models deployed with explicit action protocols, prospective calibration monitoring, and local recalibration. Future work should focus on pragmatic randomised implementation trials with hard renal endpoints, including in South-East Asian populations.