Kidney disease is a silent but formidable global health challenge. Chronic kidney disease (CKD) affects an estimated 850 million people worldwide, while acute kidney injury (AKI) contributes to millions of hospitalizations each year. Despite decades of medical progress, many patients are diagnosed too late, stratified imprecisely, or receive care shaped by tools designed before the era of big data. That era is now changing rapidly.
Artificial intelligence (AI) and machine learning (ML) technologies have entered nephrology with transformative potential. From predicting AKI hours before it becomes clinically apparent, to reading kidney biopsy slides with the accuracy of an expert pathologist, these technologies are reshaping what is possible in the diagnosis and management of kidney disease. The International Society of Nephrology has recognized AI integration as a strategic priority for the future of nephrology practice. The tools emerging from clinical research are not merely incremental improvements — they represent a fundamental shift in how kidney disease may be detected, tracked, and treated in the years ahead.
This article explores the current landscape of AI in nephrology, the clinical evidence behind key applications, the equity considerations that must guide implementation, and the role that organization ISN (International Society of Nephrology) plays in ensuring these technologies serve all patients fairly and effectively.
Understanding the Problem: Why Nephrology Needs Smarter Tools
The kidney is a remarkably complex organ, responsible for filtering waste, regulating blood pressure, maintaining electrolyte balance, and supporting red blood cell production. When the kidney fails — whether acutely or over years — the consequences ripple across virtually every organ system.
Clinicians in nephrology face a distinctive diagnostic challenge: the kidney’s reserve capacity is substantial, meaning significant damage can accumulate before conventional markers such as serum creatinine or estimated glomerular filtration rate (eGFR) clearly reflect it. By the time a patient’s numbers shift markedly, the window for early intervention may already be narrowing.
Traditional risk stratification tools, while valuable, are typically based on a small number of clinical variables, were derived from specific study populations, and cannot easily process the enormous volumes of longitudinal data that modern electronic health records (EHRs) now contain. This is precisely where AI offers something qualitatively new: the ability to learn patterns from thousands of variables across millions of patient records, detecting signals invisible to conventional statistical analysis.
Predicting Acute Kidney Injury Before It Happens
Among the most clinically compelling AI applications in nephrology is the early prediction of acute kidney injury. AKI — a sudden episode of kidney failure or damage — is common in hospitalized patients and is associated with significantly increased mortality, prolonged hospital stays, and elevated long-term risk of CKD progression. By the time AKI is formally diagnosed using conventional criteria, the damage is already underway and treatment options are narrowed.
A landmark study by Tomašev and colleagues at DeepMind, published in Nature in 2019, demonstrated what AI can accomplish in this domain. The researchers trained a deep learning model on electronic health records from more than 700,000 patients in the United States Veterans Affairs healthcare system. The model predicted AKI onset up to 48 hours before clinical diagnosis, achieving an area under the curve (AUC) of 0.92 — a metric of discriminative accuracy where 1.0 is perfect prediction and 0.5 is chance. This substantially outperformed conventional risk scoring tools in clinical use.
A 48-hour window is clinically meaningful: it allows time to adjust nephrotoxic medications, optimize fluid management, increase monitoring frequency, or involve nephrology teams proactively. Early intervention can reduce AKI severity and, in some cases, prevent permanent damage that might otherwise accelerate CKD progression. Rigorous external validation across diverse patient populations remains essential before such models enter routine clinical workflows — a standard that the International Society of Nephrology has consistently emphasized.
AI in Renal Pathology: Teaching Computers to Read Biopsies
Kidney biopsy remains the gold standard for diagnosing many forms of glomerular disease — conditions affecting the kidney’s microscopic filtering units. But biopsy interpretation is complex, time-consuming, and subject to variability between pathologists, particularly in multicenter studies. This inter-observer variability can affect treatment decisions and complicate clinical trial interpretation.
Deep learning algorithms applied to digitized biopsy specimens are beginning to address this challenge. Convolutional neural networks (CNNs) — AI models well-suited to image recognition — can be trained to identify and classify microscopic patterns in tissue samples with remarkable precision, learning from thousands of labeled images in a process analogous to how a pathologist builds expertise over years of practice.
A study published in the Journal of the American Society of Nephrology in 2021 demonstrated that a CNN achieved pathologist-level accuracy in classifying IgA nephropathy specimens according to the Oxford MEST-C scoring system. IgA nephropathy is the most common primary glomerulonephritis worldwide, and the Oxford score guides prognosis and treatment decisions. AI-assisted classification could serve as a quality control tool and enhance consistency across institutions — particularly in settings where renal pathology expertise is limited.
What This Means for Patients
Automated biopsy analysis does not replace pathologists — it can serve as a second opinion, a quality control mechanism, and a means of extending expert-level analysis to settings where specialist renal pathologists are scarce. In many regions where the ISN operates through its global programs, access to nephropathology expertise is limited. AI-assisted pathology could help reduce the diagnostic gap between high-resource and lower-resource settings — one of the ISN’s core strategic concerns.
Future directions in this field include:
- Automated quantification of interstitial fibrosis and tubular atrophy — key histological markers of CKD severity that currently require subjective estimation
- Integration of pathology images with clinical and genomic data for comprehensive, individualized disease profiling
- Development of AI tools validated across diverse biopsy preparation techniques, staining protocols, and digital scanner types
- Multi-disease models capable of classifying a range of glomerular conditions beyond IgA nephropathy
Predicting CKD Progression: From Population Models to Precision Risk
Chronic kidney disease progresses at highly variable rates. Some patients with stage 3 CKD remain stable for decades; others deteriorate to end-stage renal disease (ESRD) within a few years. Identifying which patients are at highest risk is essential for treatment planning, specialist referral timing, and patient counseling.
The Kidney Failure Risk Equation (KFRE), developed by Tangri and colleagues and published in JAMA in 2011, represents a major advance in this area. Using four to eight clinical variables — including eGFR, urinary albumin-to-creatinine ratio, age, and sex — the KFRE predicts a patient’s two- and five-year risk of kidney failure with high discriminative accuracy. It has been externally validated in more than 30 countries, demonstrating robust performance across diverse healthcare systems. The International Society of Nephrology has endorsed the KFRE in clinical guidelines as a practical risk stratification tool, informing decisions about nephrology referral timing, dialysis preparation, and transplant evaluation.
Next-Generation Predictive Models
Beyond the KFRE, more sophisticated machine learning models are being developed that incorporate additional data streams to further refine risk prediction:
| Data Type | Examples | Potential Contribution |
|---|---|---|
| Clinical variables | eGFR trajectory, blood pressure control, comorbidities | Captures disease severity and treatment response over time |
| Biomarker trajectories | Proteinuria trends, serum bicarbonate patterns | Reflects dynamic changes in kidney function and injury |
| Genomic data | Polygenic risk scores, rare variant analysis | Identifies heritable susceptibility and underlying disease mechanisms |
| Imaging data | Kidney volume on MRI, cortical thickness measurements | Structural markers that correlate with functional progression |
| EHR-derived signals | Medication adherence, outpatient visit patterns | Behavioral and socioeconomic context influencing outcomes |
These multi-modal models hold considerable promise, but also introduce new challenges: greater complexity, higher data requirements, reduced interpretability, and more opportunities for bias to enter the training pipeline. The nephrology community must engage with these challenges rigorously, rather than allowing enthusiasm for technical innovation to outpace critical evaluation.
The Equity Imperative: AI Must Work for Everyone
Perhaps the most critical dimension of AI in nephrology — and the one where the International Society of Nephrology has spoken most clearly — is algorithmic equity. AI models learn from historical data. If that data reflects existing disparities in healthcare access, diagnosis, or treatment, the resulting model may perpetuate or amplify those disparities, embedding systemic inequity into the digital infrastructure of future care.
The issue of race-based clinical algorithms in nephrology has received substantial and overdue attention. For decades, eGFR equations used a race coefficient that systematically estimated higher kidney function in Black patients compared to patients of other races with the same serum creatinine level. This coefficient led to underestimation of kidney disease severity in Black patients, delayed referrals to nephrology, and reduced access to transplant evaluation. The harms were real and measurable.
When AI models were trained on datasets that included race-based eGFR values, they inherited and encoded this inequity. A commentary published in the Journal of the American Society of Nephrology in 2021 documented how race-based eGFR equations embedded in prior AI training pipelines perpetuated health disparities in automated clinical decision-making. In response, ISN (International Society of Nephrology) supported the transition to the race-free CKD-EPI 2021 equations — updated formulas that eliminate the race coefficient entirely. This shift reflects the ISN’s broader commitment to algorithmic equity in kidney care.
Principles for Equitable AI Development in Nephrology
- Diverse training datasets: Models should be developed using data that adequately represents patients across racial, ethnic, geographic, and socioeconomic groups — not simply reflecting the populations most conveniently available to researchers.
- Prospective validation in underrepresented populations: External validation studies must specifically assess model performance across demographic subgroups, not only in aggregate metrics that can mask differential performance.
- Transparent and stratified reporting: Publications describing AI nephrology tools should report performance metrics stratified by relevant demographic and clinical variables, enabling the community to identify disparities.
- Ongoing post-deployment monitoring: AI tools in clinical use should be systematically monitored for differential performance across patient groups over time, with clear mechanisms for correction when disparities are detected.
- Community engagement: Patients and communities disproportionately affected by kidney disease should have a meaningful voice in the design, governance, and evaluation of AI systems used in their care.
AI in Dialysis: Supporting the Most Vulnerable Patients
Patients receiving dialysis represent some of the most medically complex individuals in medicine. Machine learning models are being investigated to predict intradialytic hypotension — a dangerous blood pressure drop during hemodialysis sessions — allowing clinicians to adjust treatment parameters proactively. Outcome prediction tools may also support individualized discussions about transplant candidacy, palliative care, and modality selection. As with all AI applications in nephrology, rigorous validation across diverse populations remains essential before clinical deployment.
Barriers to Implementation and the Path Forward
Despite impressive research results, translating AI tools into routine clinical practice involves significant challenges the nephrology community must address.
- Data infrastructure: AI models require large, well-structured datasets. Many healthcare systems, particularly in low- and middle-income countries, lack the digital infrastructure and data governance frameworks needed to support this.
- Model interpretability: Many high-performing AI models function as “black boxes,” offering predictions without understandable explanations. This opacity limits clinical trust and adoption.
- Regulatory pathways: AI-based decision support tools must navigate approval processes that vary by country and are still evolving to address the specific characteristics of adaptive machine learning systems.
- Clinical workflow integration: Even a well-validated model will not impact care if it cannot be embedded in clinical workflows in a way that is actionable and free from alert fatigue.
- Model drift: Patient populations and clinical practices evolve. AI models may degrade in performance over time if not periodically retrained — requiring sustained governance investment.
The ISN is actively engaged in addressing these barriers through global education programs, clinical practice guideline development, and advocacy for evidence-based policy in kidney health. The ISN’s position is clear: AI must be developed responsibly, validated rigorously, and implemented equitably — or its potential benefits will remain out of reach for the patients who need them most.
The Role of the International Society of Nephrology in Shaping AI’s Future
As the leading global organization in nephrology, the International Society of Nephrology occupies a uniquely important position in guiding how AI develops within the field. The ISN brings together nephrologists, researchers, patient advocates, and policymakers from across the world — providing a platform for setting standards, sharing evidence, and advancing policy in kidney health.
The ISN’s endorsement of the KFRE, its support for race-free eGFR equations, and its identification of AI integration as a strategic priority all reflect a coherent vision: that technological innovation in nephrology must be grounded in scientific rigor, inclusive of diverse populations, and oriented toward improving outcomes for all patients — not only those in high-resource academic medical centers.
Looking ahead, the ISN is well
positioned to lead in several areas:
- Developing consensus guidance on validation standards and reporting requirements for AI tools in nephrology
- Advocating for meaningful inclusion of underrepresented populations in AI training and validation datasets
- Supporting education programs to help practicing nephrologists critically evaluate and appropriately apply AI-based clinical tools
- Facilitating international collaborations to build the diverse datasets needed for globally applicable AI models
- Engaging regulatory bodies to develop frameworks enabling responsible AI adoption in kidney care worldwide
Conclusion: A Technology With Promise, and a Responsibility to Match
Artificial intelligence is not a distant prospect for nephrology — it is already here, embedded in research pipelines, entering clinical pilots, and beginning to influence how kidney disease is diagnosed and managed. The evidence is compelling: AI can predict AKI earlier than conventional tools, classify renal biopsies with expert-level accuracy, stratify CKD progression risk across diverse populations, and support more individualized dialysis care. These are meaningful advances where earlier, more accurate clinical action can make the difference between preserved kidney function and irreversible failure.
Yet the promise of AI is matched by an inseparable responsibility to deploy it carefully and equitably. Algorithms trained on biased data will produce biased outcomes. Tools validated in one population may perform differently in another. Models that clinicians cannot understand will not be used effectively, regardless of technical performance on benchmark datasets. These are not hypothetical concerns — they are lessons already learned, in some cases at real cost to patient wellbeing.
The International Society of Nephrology stands at the center of this pivotal moment, with the credibility to set meaningful standards and the global reach to ensure those standards are adopted broadly. As the ISN continues to shape the integration of AI into nephrology practice, the guiding principle must remain constant: that innovation serves patients — all patients, everywhere, equitably. That is the standard to which the nephrology community, and every AI tool it adopts, must be held.