From Technical Performance to Clinical Impact: Challenges and Opportunities for AI Modelling of CKD Progression
Background. Chronic kidney disease (CKD) is a major global health burden, with highly heterogeneous trajectories ranging from long-term stability to rapid progression towards kidney failure and kidney replacement therapy. Accurate prediction of CKD progression is essential to support personalised monitoring, timely therapeutic intensification, nephrology referral, and preparation for advanced kidney care. Artificial intelligence (AI), machine learning (ML), and deep learning (DL) approaches may improve risk stratification by modelling complex, nonlinear, and time-dependent patterns in routinely collected clinical data, electronic health records, and potentially multimodal data sources. Nonetheless, it remains unclear how ready these AI-driven models are for clinical use. This structured narrative review aimed to map the current landscape of AI for predicting CKD progression, assess the robustness and clinical relevance of the evidence, and identify methodological and implementation gaps that must be addressed to translate technical accuracy into clinical impact. Summary. This review synthesised 35 studies published between 2020 and 2025 that developed and evaluated models to predict CKD progression. Most studies used retrospective cohort designs and electronic health record data, relying mainly on demographic variables, laboratory measures, comorbidities, and clinical history. Tree-based models, random forests, boosting methods, regression-based approaches, and neural networks were the most commonly investigated methods. Overall, the reviewed models often achieved favourable discriminative performance, as measured by AUROC or C-index. Nevertheless, the evidence base remains limited by substantial heterogeneity in study designs, cohort compositions, prediction targets, time horizons, feature engineering, and validation strategies. External validation was reported in only about one-quarter of studies, while calibration, clinical utility, patient-level uncertainty, and implementation-oriented evaluation were inconsistently addressed. Prediction endpoints were predominantly kidney failure-related. Multimodal data, including imaging, genomics, and social determinants of health, were rarely used, and explainability methods were more often applied for model description than for clinician-facing decision support. Key Messages. AI models show substantial promise for predicting CKD progression, particularly when trained on routinely available clinical data. However, the field remains more advanced in model development than in clinical validation and implementation. Future research should prioritise external and temporal validation, systematic calibration assessment, decision-curve or net-benefit analysis, comparison with established clinical risk scores, clinically meaningful prediction horizons, subgroup and fairness evaluation, patient-level uncertainty estimation, and prospective testing of model-informed care. These steps are essential to move from statistically accurate models to reliable, interpretable, equitable, and actionable tools for CKD management.
Authors
- Anna Caroli (ORCID: https://orcid.org/0000-0002-4130-4663)
- Andrea Remuzzi (ORCID: https://orcid.org/0000-0002-4301-8927)
- Matteo Suardi
- Alberto Arrigoni
Publication Details
- Journal
- The Nephron journals/Nephron journals
- Published
- 2026-11-21
- DOI
- https://doi.org/10.1159/nef/adtag022
- Primary Topic
- Machine Learning in Healthcare
- Type
- article
- Field-Weighted Citation Impact
- 0.00