AI in kidney care shows promise but clinical proof still lags behind

Artificial intelligence is spreading rapidly through kidney medicine, with models now reported for predicting acute kidney injury, tracking chronic kidney disease progression, reading biopsy slides, forecasting transplant outcomes, and fine-tuning dialysis. A narrative review published in Frontiers in Nephrology on 30 October 2026 examined this expanding literature and reached a sobering conclusion: many models show impressive discrimination on paper, but evidence of real clinical utility remains scarce. The review, led by Moses Ishaku Madami and colleagues at Shanxi Medical University, synthesised studies published mainly between 2020 and 2025 and evaluated them for validation, interpretability, fairness, and readiness for actual deployment.

The authors searched PubMed on 24 November 2025 using terms spanning artificial intelligence, machine learning, deep learning, neural networks, natural language processing, large language models, generative AI, computer vision, reinforcement learning, and federated learning, combined with nephrology and kidney-disease terms. The search identified 9,179 records, of which 3,617 survived title and abstract screening for full-text assessment. Screening was performed independently by the first and second authors, with disagreements resolved by a third. Two foundational pre-2020 studies were added through citation chaining because of their importance to the field’s development. Because this was a narrative review rather than a quantitative systematic review, the authors selected studies most relevant to predefined clinical domains for detailed synthesis, and they deliberately distinguished between model development, internal validation, external validation, prospective clinical evaluation, and demonstrated clinical utility.

The sharpest signal came from acute kidney injury, the field’s flagship application. A recurrent neural network trained on 703,782 Department of Veterans Affairs patients predicted 55.8 percent of inpatient AKI episodes up to 48 hours before clinical recognition, achieving an area under the receiver operating characteristic curve of 92.1 percent for any AKI stage. Successive multinational studies extended this work, with the largest external validation outside the VA ecosystem covering 16,760 intensive care patients across 145 units in the United States, the Netherlands, and Italy, reaching an AUROC of 0.884 for KDIGO stage 2 or 3 AKI with more than 14 hours of lead time. An interpretable tree regressor for persistent stage 3 AKI was validated in two independent multicentre cohorts, and a biomarker-based XGBoost model called AKIMLpred, integrating nine urinary kidney injury biomarkers including NGAL, KIM-1, IGFBP7, IL-18, L-FABP, cystatin C, CAF22, and NT-proBNP, achieved an AUC of 0.902 in an external test cohort with an accuracy of 0.846, outperforming models built on routine clinical variables alone.

Yet the review repeatedly flagged the same caveat. The original VA cohort was 93.6 percent male, and models trained on it underperformed in female populations. The authors documented reproducibility failures across externally validated AKI models, noting that impressive development-cohort AUROCs must not be extrapolated uncritically. A prospective multicentre paediatric ICU study translated a SHAP-interpreted random forest model into a real-time clinical web application with demonstrated feasibility and usability, but reduced dialysis dependence and improved mortality were never measured. The review described that outcome evaluation as the field’s central unfinished task. A point-of-care model using only features available at emergency presentation reached an AUC of 0.85 with a recall of 0.63 and a precision of 0.54, sidestepping the baseline creatinine problem that complicates admission-stage deployment, but this too stops short of proving that AI-assisted care changes patient outcomes.

In chronic kidney disease, the authors judged AI models against an established benchmark rather than in isolation. Liu and colleagues evaluated and recalibrated the Kidney Failure Risk Equation in a multinational, population-based cohort, providing a reference point for newer prediction models. Machine-learning approaches have since explored additional predictors, personalised progression prediction, TabNet and LSTM-based detection, and multimodal fusion models that may capture nonlinear relationships and longitudinal patterns not fully represented by conventional risk equations. However, the review concluded that higher discrimination alone does not establish clinically meaningful incremental value over the KFRE. Evidence of improved calibration, reclassification, decision-analytic performance, or patient outcomes would be needed before replacing or routinely supplementing established tools. Current evidence therefore supports AI as a promising complement to conventional CKD risk prediction rather than establishing its superiority for clinical decision-making.

Nephropathology and transplantation showed some of the strongest multicohort evidence. A multicentre proof-of-concept study applied deep learning to kidney transplant pathology, classifying biopsies into normal, rejection, and other categories with reported AUC values of 0.80 to 0.87. Other work explored automated quantification of IgA nephropathy histological features and deep-learning quantification of interstitial fibrosis in digitised biopsies, while ultrastructural assessment of glomerular morphology remains comparatively early. In transplantation, the iBox system integrated clinical and transplant-related information to predict long-term allograft failure and was compared with predictions by transplant physicians. Deep-learning approaches evaluated for allograft histology included automated classification based on morphological and molecular Banff features, with one multiple-instance learning model achieving an AUC of 0.798 for rejection classification and 0.936 for one-year graft-loss prognosis in independent testing. Dynamic AI-based prediction of renal survival was evaluated in an international multicohort observational study with an AUC of 0.857. The review stressed that these studies indicate substantial progress in external and multicohort evaluation, but predictive or diagnostic performance should not be interpreted as evidence that AI improves transplant decision-making or graft outcomes, and prospective interventional evaluation remains necessary before claims of routine clinical deployment can be justified.

Dialysis applications revealed an even wider gap between technical performance and clinical proof. A recurrent neural network evaluated across 261,647 haemodialysis sessions achieved an AUROC of 0.94 for one definition of intradialytic hypotension, with additional AUROCs of 0.87 and 0.79 for alternative definitions. The review characterised this as strong retrospective predictive performance that does not establish prospective clinical utility or improved patient outcomes. Reinforcement learning has been applied to dry-weight assessment using a dueling double deep Q-network, framing fluid optimisation as a sequential decision-making problem, but this remains simulation-based and unvalidated prospectively. In peritoneal dialysis, prediction models for peritonitis reported individual AUCs ranging from 0.659 to 0.997, although a systematic review and meta-analysis identified substantial methodological limitations and a high risk of bias across the available models. A random-forest model for technique failure after peritoneal dialysis-associated peritonitis reported an AUC of 0.916 with accuracy of 93.70 percent, sensitivity of 96.67 percent, and specificity of 86.49 percent in a retrospective cohort of 508 episodes. The authors judged these results promising in discrimination but not evidence of external validity or clinical benefit, noting that the limited number of multicentre and prospective studies remains a major barrier to translation.

Emerging technologies received a more cautious assessment. Large language models were evaluated on a nephrology benchmark of 975 questions, where GPT-4 answered 74 percent correctly and GPT-3.5 answered 51 percent, but the review framed this as knowledge benchmark performance rather than clinical utility or patient benefit. In a separate randomised trial of 50 physicians, GPT-4-assisted diagnostic reasoning reached 76 percent on a prospective RCT design but did not significantly improve physician reasoning compared with conventional resources, and the study was not nephrology-specific. Federated learning, which allows collaborative model development without centralising patient-level data, was represented by a non-nephrology study of 6,314 patients across 71 sites on six continents for tumour-boundary detection, included as a methodological reference rather than clinical evidence. The review concluded that nephrology-specific evidence for both LLMs and federated learning remains comparatively limited, with benchmark performance not establishing reliability, safety, or patient benefit.

Across all domains, the authors applied a qualitative evidence-maturity and translational-readiness framework, explicitly stating that it is not a validated numerical scoring instrument and does not assign thresholds for deployment. AKI prediction and nephropathology rated moderate to high in evidence maturity, CKD progression moderate, haemodialysis moderate, peritoneal dialysis early to moderate, and large language models, reinforcement learning, and federated learning early to early-moderate. Translational readiness was rated emerging for the more mature domains and early for the rest. Recurring challenges included inconsistent calibration reporting, generalizability concerns across geographic and demographic settings, interpretability gaps, algorithmic fairness, data governance, and integration into clinical workflows. The authors noted that fairness-blind CKD risk tools may silently delay specialist referral for Black and Hispanic patients, and that large language models can propagate discredited race-corrected kidney function estimates into clinical consultations.

The review proposed six research priorities: prospective trials with standardised AKI, CKD, and dialysis endpoints; an international multicentre annotated whole-slide-image registry for nephropathology AI trials in lupus nephritis and IgA nephropathy; standardised peritoneal dialysis data capture and validation studies; federated governance frameworks adapted from the Kidney Precision Medicine Project; human-AI integration research using randomised simulated trials and human-factors engineering; and mandated demographic diversity reporting with LMIC-specific development and validation consortia. The authors acknowledged their own limitations, including the rapid evolution of the field, the relative scarcity of peritoneal dialysis evidence, and the subjectivity inherent in their qualitative framework. Their bottom line was consistent throughout: AI has substantial potential to support precision nephrology, but technical performance should not be equated with clinical effectiveness or deployment readiness, and AI should complement rather than replace clinical expertise, with its ultimate value determined by demonstrated benefit within real-world nephrology practice.

Subject of Research: A narrative review of artificial intelligence applications across nephrology domains and their translational readiness.

Article Title: Artificial intelligence for precision nephrology: current evidence to clinical translation and future prospects

Article References: Madami, M. I., Amaoge, C. S., Wulandari, A., & Zhou, X. (2026). Artificial intelligence for precision nephrology: current evidence to clinical translation and future prospects. Frontiers in Nephrology, 6, Article 1950125. https://doi.org/10.3389/fneph.2026.1950125

Image Credits: AI Generated

DOI: 10.3389/fneph.2026.1950125

Keywords: artificial intelligence, nephrology, acute kidney injury, chronic kidney disease, machine learning, kidney transplantation, haemodialysis, peritoneal dialysis, large language models, federated learning, clinical validation, precision medicine