AI in Medicine & Nephrology

It helps to say plainly what AI is good at today, because the honest version is more useful than the marketing version. AI is good at watching many variables at once. It is good at noticing risk earlier than a busy team can. It is good at measuring things on images and pathology slides that are tedious to count by hand. It is good at summarizing a messy hospital course into something a consultant can read in a minute. And it is good at standardizing the repetitive decisions that should be the same every time but often are not.

Notice what is not on that list. AI does not yet replace clinical judgment. It does not weigh tradeoffs the way a chief does at the bedside. And it does not own the consequences of a treatment decision. Those still belong to the physician. So the right picture is not a machine that takes your place. It is a machine that handles the parts of the hospital that currently force a good mind to hunt through fragmented data, delayed signals, and incomplete histories.

Why now, and not ten years ago? Two things changed. Hospitals went fully electronic, so a patient's labs, vitals, medications, notes, scans, and pathology now live as data a computer can actually read. And the methods for finding patterns in that data got dramatically better. Put those together and you get software that can read the whole chart continuously, across time, instead of firing a single alert when one number crosses one line.

This site follows the arc you would expect from a teacher. First, plain definitions, so no term later is mysterious. Then AI across medicine broadly, where the most mature uses are imaging and pathology, not chatbots. Then nephrology. Then onco-nephrology, the heart of it, written for the person who ran a kidney service inside a cancer hospital. The closer we get to your field, the more concrete it becomes, because that is where this technology and your career meet.

One promise about tone. On the medicine, you will be spoken to as a peer, because you know it cold. On the technology, every new idea gets explained in plain English with a clinical example the first time it appears. The goal is not to dazzle you. The goal is to hand you something accurate enough that you could pass it to a colleague.

The simplest way to think about all of this: "AI" is an umbrella word for software that finds patterns in data and uses them to make a prediction, read an image, or produce text. It is not one thing. In medicine it usually means one of five things, and they are easy to keep straight once you see a kidney example for each.

The most important idea to carry forward is a shift in how this software works. The old generation of clinical decision support was rule-based. A human wrote a fixed rule, the computer obeyed it, and that was the whole of its intelligence. The classic example is the creatinine alert: "if creatinine rises by X, alert the team." Useful, but blunt. It looks at one number against one threshold, and it cannot tell a dangerous rise from a trivial one, because it does not understand context.

Modern AI works differently. It reasons across time and across many kinds of data at once. Instead of watching a single lab against a single line, it can notice that a patient's creatinine, magnesium, albumin, cisplatin dose, neutrophil trend, recent contrast exposure, and proton pump inhibitor use, taken together, imply a higher-risk trajectory than any one of those variables would suggest alone. That is the leap: from a rule about one number to a reading of the whole picture, the way a good consultant actually thinks. The rest of this site is essentially that idea, applied to harder and harder kidney problems.

A natural first question from a clinician is: what has actually been approved? The answer is clarifying. A 2025 analysis in npj Digital Medicine catalogued 1,016 AI-enabled medical devices authorized by the FDA. Among the 736 unique devices, 84.4% work on images. Another 14.5% work on signals such as ECG or other waveforms. Less than 1% touch genomic data, and only 0.4% work on the ordinary tabular data of the electronic record. Radiology is by far the dominant specialty.

The most telling finding: as of the study's December 2024 cutoff, there were no large language model devices in the authorized set at all. A "large language model," or LLM, is the kind of AI behind ChatGPT. It is the technology getting all the public attention. And it is precisely the technology that, so far, regulators have not cleared as a medical device.

This is the useful separation: the AI that is regulated, validated, and running in clinics today is overwhelmingly the AI that looks at pictures. The AI that dominates the headlines is barely in the room yet. Hold that distinction. It makes everything below easier to judge.

Because regulated medical AI is mostly image-based, the most mature clinical applications are exactly the ones built on images and slides. The National Cancer Institute describes AI as relevant across cancer mechanisms, screening, diagnosis, drug discovery, surveillance, and care delivery. But in real-world oncology, the largest share of deployed AI sits in diagnosis, radiology, and pathology. Concretely, that means:

• Mammography, with AI flagging suspicious areas for the radiologist's attention.
• Lung nodule detection on chest CT.
• Prostate MRI interpretation.
• Polyp detection during colonoscopy, where the system marks a polyp the moment it appears on screen.
• Dermatology image classification, sorting skin lesions by appearance.
• Digital pathology, where the slide is scanned and the software helps quantify what is on it.

The common thread is augmentation, not replacement. The AI offers a read. A physician owns the decision. This is the same posture you would expect from a sharp fellow: a second set of eyes that never tires, presenting findings for your judgment.

Radiation oncology offers one of the cleanest, most concrete examples of AI earning its place. Before a patient is treated, someone has to draw, slice by slice on the CT scan, the exact outline of the tumor and the exact outlines of the healthy organs nearby that must be spared. The kidneys, the spinal cord, the salivary glands, the optic nerves. This careful tracing is called contouring, or segmentation. It is skilled, it is repetitive, and it takes a long time.

"Auto-contouring" means an AI, trained on thousands of prior cases, draws those outlines first. The physician then reviews and corrects them rather than starting from a blank scan. A Mayo Clinic study, working with Google Health, trained a deep-learning model to auto-contour head and neck cases covering 42 organs at risk. It cut contouring time by 76%, and randomized review by radiation oncologists confirmed the time was saved without compromising contour accuracy.

Matching a patient to the right clinical trial is harder than it sounds, and the reason is text. Trial eligibility criteria are long, intricate, and written in prose. A trial might exclude patients with a certain prior therapy, a certain lab value, a certain mutation, or a degree of kidney impairment. The patient's own relevant facts are scattered across notes, scans, and reports in equally messy language. Manually checking one patient against one trial's criteria, then repeating for dozens of trials, is slow and error-prone. Good matches get missed simply because no one had time to read everything.

TrialGPT, published in Nature Communications in 2024, used large language models to do this reading. It worked in three stages: retrieving candidate trials, checking the patient against each eligibility criterion, and ranking the trials by fit. Its rankings correlated strongly with the judgments of human experts, and in a user study it reduced screening time by 42.6%. The point is not that AI decides who enters a trial. A research nurse or oncologist still makes that call. The point is that AI can read all the fine print on every patient, consistently, and hand the clinician a short, sensible list to confirm.

If one development deserves the word "future," it is the AI prediction of protein structure. Here is why it matters, in plain terms. Proteins are the tiny machines that run the body. Each one folds into a specific three-dimensional shape, and that shape is what determines its job and how a drug can act on it. A drug molecule works by physically fitting against a protein, the way a key fits a lock. So if you want to design a drug, you badly want to know the exact shape of the protein you are targeting.

For decades, working out a single protein's shape in the laboratory could take a scientist months or years. AlphaFold, built by DeepMind, is an AI that predicts those shapes from the protein's sequence in minutes, at an accuracy that stunned the field. In roughly sixty years, scientists had determined about 150,000 protein structures by hand. AlphaFold has now predicted the structures of over 200 million proteins.

For medicine, this collapses one of the slowest steps in drug discovery. Researchers can see the shape of a target, understand how molecules might bind to it, and design candidate drugs far faster than before. The film below tells that story better than any text could.

AI is being applied across the rest of the drug pipeline too: target discovery, screening large libraries of compounds, biomarker selection, and toxicity prediction, which means forecasting whether a candidate drug is likely to harm the body before it is ever given to a person.

That last category reaches directly into your field. Researchers are building AI and machine-learning models to predict drug-induced kidney injury from chemical, biological, and clinical data. In time, this could help flag, before a drug is approved or prescribed, which compounds are most likely to cause tubular injury, electrolyte disturbance, or progressive CKD. The renal toxicity that you have spent a career managing after the fact may increasingly be predicted before it happens.

The last category is the LLMs from the opening section, now applied to the part of medicine no clinician misses: paperwork. Large language models are being tested to draft clinical notes, summarize long hospitalizations before a consult, extract a creatinine baseline and an AKI timeline from a tangled chart, list a patient's nephrotoxic exposures, translate instructions into plain language for patients, and prepare cases for tumor board or nephrology conference. A 2025 study compared notes generated by ambient AI, software that listens to a visit and drafts the note, against physician-written notes across five specialties. The technology is moving into real workflows quickly.

But the honest caution belongs here, and a chief will want it stated plainly. A 2026 study in JAMA Network Open tested frontier language models on standardized clinical vignettes. The models did reasonably well at naming a final diagnosis or a management step, with failure rates usually under 0.40. They did poorly at the part that matters most to a diagnostician: building a differential. Failure rates for generating a differential diagnosis exceeded 0.80. The models tended to collapse prematurely onto a single answer instead of holding several possibilities open and refining between them, which is exactly the discipline a clinician uses when facing uncertainty.

The lesson is not that the tools are useless. It is that they are drafting tools and summarizing tools, to be used under a physician's review, not unsupervised reasoners that can be trusted with a differential or with what they do not know.

The flagship example came in 2019, from DeepMind (Google's AI lab) working with the U.S. Veterans Affairs health system, published in Nature. They trained a model on a very large set of patient records to predict AKI before it was clinically obvious. This was "deep learning," which means the program was not given rules like "alert when creatinine rises by X." Instead it was shown the records of hundreds of thousands of patients and left to find, on its own, the combinations of labs, vitals, medications, and trends over time that tend to precede injury. Because it works across time, it can weigh a trajectory, not just a single value.

The result. The model predicted 55.8% of inpatient AKI episodes and 90.2% of the AKI episodes severe enough to later require dialysis, up to 48 hours in advance. That second number is the striking one. The catch is that it also cried wolf: roughly two false positives for every true positive.

The same kind of model is being pointed at chronic kidney disease, where the question is not "will this patient be injured tonight" but "which of these patients is going to decline, and how soon." Fed the usual inputs, labs, demographics, comorbidities, medication history, the model can rank a population by risk.

The value here is organizational. Imagine a health system that can look across every patient with stage 3 or 4 CKD and surface the ones most likely to progress. Those are the patients you would want to move to the front of the line: a nephrology visit sooner, an evaluation for SGLT2 inhibitors, blood-pressure optimization, a medication review for nephrotoxins, transplant education while there is still time, vascular-access planning before it becomes urgent. The model does not decide any of this. It decides who gets your attention first.

A 2026 study in JAMIA on predicting end-stage renal disease outcomes made the point that matters most for this to be trustworthy. The models that hold up pull from many sources of data at once, they can show why they flagged a given patient rather than just producing a number, and they are built to guard against bias. That last point is not a footnote. CKD and ESRD outcomes are shaped heavily by race, access, and referral patterns, and a model trained on biased care will repeat it unless someone designs against that.

Dialysis may be the most natural fit of all, because it throws off so much data: blood pressure, ultrafiltration rate, blood flow, dialysate, symptoms, interruptions, medication timing, lab trends, session after session. That density is exactly what these models feed on.

The current focus is predicting trouble mid-session. A study in Scientific Reports built an "explainable" deep-learning model to predict both intradialytic hypotension and intradialytic hypertension during the session, in time to act. "Explainable" is the operative word: rather than just sounding an alarm, the model can show which factors drove the warning, which is the difference between something a clinician trusts and something a clinician learns to ignore.

The broader idea has been called "precision dialysis," using this volume of data to personalize the prescription rather than run everyone on the same settings. The near-term uses are unglamorous and real: safer ultrafiltration targets, better dry-weight estimation, earlier nursing intervention before a patient crashes, detection of access problems, and even the operational side, staffing and chair scheduling.

In transplant, AI is being studied across the full arc: donor-recipient matching, graft survival, waitlist mortality, allocation policy, rejection risk, and biopsy interpretation. The recurring caveat in this literature is fairness and explainability, for the same reason as in CKD: allocation decisions carry equity consequences, so a model that cannot show its reasoning is hard to defend.

The most striking single result is a "virtual biopsy" system published in Nature Communications in 2024. The team assembled 14,032 day-zero biopsies from 17 centers and trained a model to estimate, from just 11 basic donor parameters, what that biopsy would likely show: the Banff lesions and the percentage of sclerotic glomeruli. They reported good discrimination, calibration, and robustness, and released a ready-to-use online tool for clinicians.

This does not retire the biopsy. Day-zero biopsies are invasive, costly, and can delay a transplant, and many centers do not perform them. The promise is a reliable estimate when a real biopsy is not available, and better judgment about which biopsies are most worth doing.

Digital pathology is the clearest case where AI augments the nephrologist's eye. The biopsy is information-rich, but much of the work is laborious and subject to interobserver variability. A model can count and measure at scale, consistently. Deep-learning models have been built to quantify glomerulosclerosis on whole-slide biopsy images, and to estimate interstitial fibrosis and tubular atrophy, in some work even from ultrasound rather than tissue.

The key idea for you is that this is not only about diagnosis. It makes pathology quantitative and reproducible: a percent sclerosis, a fibrosis burden, a lesion map, a score that means the same thing across readers and can be compared to a cohort. Quantitative and reproducible is the door to prognostic, tying histologic pattern to outcome in a way that a qualitative read never quite could.

A single oncology patient may carry metastatic disease, CKD, recent cisplatin, an immune checkpoint inhibitor, IV contrast from staging scans, broad antibiotics, a proton pump inhibitor, hypoalbuminemia, magnesium wasting, intermittent sepsis, and a creeping obstruction. Any one of those explains a rising creatinine. Several of them together change the answer.

A 2018 PLOS One study addressed AKI prediction in cancer patients specifically, and the phrase its authors used matters: heterogeneous and irregular clinical data. Cancer care does not generate clean, evenly spaced numbers. Patients cycle through infusion visits, admissions, surgeries, scans, antibiotics, and changing regimens, so the data arrive in bursts and gaps. The interesting part of that work was building something that tolerates the mess your patients actually produce.

The application that follows is a cancer-center AKI risk dashboard. Picture a single screen that watches creatinine slope, eGFR, urine findings, volume proxies, medication exposures, chemotherapy timing, contrast, sepsis markers, and obstruction risk. Its output is not the word "AKI." It is a ranked differential: cisplatin toxicity, ICI nephritis, prerenal azotemia, ATN, obstruction, tumor lysis, myeloma-related disease, drug-induced AIN. The same list you would write, assembled before the consult.

This is the most concrete example in the chapter, because it is a clean, externally validated risk score rather than a black box. A 2024 BMJ cohort study (open-access full text) derived and externally validated a simple risk score for severe AKI after a first dose of intravenous cisplatin, in adults treated between 2006 and 2022. The derivation cohort held 11,766 patients and the external validation cohort held 12,951. Severe cisplatin-associated AKI occurred in 5.2% of the derivation cohort and 3.3% of the validation cohort. The predictors will look familiar to you: age, hypertension, diabetes, baseline creatinine, hemoglobin, white blood cell count, platelets, albumin, magnesium, and cisplatin dose.

The model's C-statistic was 0.75, which is its ability to rank a higher-risk patient above a lower-risk one, where 0.5 is a coin flip and 1.0 is perfect. Earlier models sat around 0.60 to 0.68. The plain way to describe what this does: it turns decades of clinical intuition into a consistent, risk-stratified workflow, then applies it before the injury rather than after. Before the infusion, the system estimates severe AKI risk and lets the team standardize hydration and magnesium, route high-risk patients to nephrology, weigh alternatives where appropriate, and set sensible post-infusion lab timing.

The checkpoint inhibitors opened a new renal complication space, and it is exactly the kind of problem where ranking beats alerting. The American Society of Onco-Nephrology position statement frames the spectrum: ICIs cause a range of kidney immune-related adverse events, most commonly acute tubulointerstitial nephritis, but also glomerular disease and electrolyte disturbances. The clinical question is never "did creatinine rise." It is whether this is ICI nephritis, ATN, prerenal AKI, obstruction, infection-related AKI, PPI-associated AIN, NSAID toxicity, or something else. That distinction drives everything that matters next: steroids, holding the cancer therapy, rechallenge, biopsy, and prognosis.

A 2024 PLOS One study used interpretable machine learning on this exact question. "Interpretable" means the model shows which factors drove its answer instead of producing a number with no reasoning attached. Working from records on 616 ICI-treated patients, it predicted AKI within seven days, then used a method called SHAP, which assigns each factor a contribution to that specific patient's risk, along with clustering, which groups similar patients together, to surface patient-specific drivers. That is the feature that earns a nephrologist's trust: not just a probability, but the why behind it.

A 2024 meta-analysis put numbers to the epidemiology. All-cause AKI occurred in 7.4% of ICI-treated patients and ICI-related AKI in 3.2%, with associations between AKI and concomitant PPIs and NSAIDs. The future use case follows directly: help decide which ICI-treated patients need urgent biopsy, which can be managed by holding nephrotoxins and rechecking labs, and which are high risk for true immune-mediated nephritis, and then support safer rechallenge decisions afterward.

Plasma-cell and monoclonal gammopathy disorders are another high-value area, because the decision is usually about whether and when to biopsy. The International Myeloma Working Group recommends a structured renal evaluation: creatinine, eGFR, free light chains, 24-hour urine protein, electrophoresis, immunofixation, and biopsy in selected cases such as significant albuminuria or lower free-light-chain levels. The Mayo MGRS prediction tool then estimates the probability that a kidney biopsy will actually show an MGRS lesion. At a probability threshold of 0.10, it was 98.9% sensitive and 50.5% specific. At a threshold of 0.25, it was 88.0% sensitive and 70.2% specific. The lower threshold is the one you would use to avoid missing a lesion. The higher one trims unnecessary procedures.

The decision this supports is the one you have made many times: do the kidney findings reflect a monoclonal process that needs hematologic treatment, or a coincidental MGUS plus unrelated CKD, or diabetic kidney disease, or amyloid, or cast nephropathy, or light-chain deposition disease. A triage tool helps avoid both underdiagnosis and an invasive procedure in a thrombocytopenic patient who did not need it.

The same logic extends across the rest of the territory you know:

• Tumor lysis syndrome: predicting which patients need aggressive prophylaxis, rasburicase, closer electrolyte monitoring, or admission.
• Obstructive nephropathy: combining imaging reports, hydronephrosis history, creatinine trajectory, tumor location, and symptoms to flag who needs ultrasound, CT review, a stent, or nephrostomy.
• VEGF inhibitor and TKI toxicity: detecting early proteinuria, hypertension, thrombotic microangiopathy signals, and risk of progressive CKD.
• CAR-T and cytokine-release syndromes: predicting AKI risk from inflammatory markers, hemodynamics, nephrotoxins, and ICU-level events.
• Pediatric oncology AKI: commonly driven by nephrotoxic drugs, but also by tumor infiltration or compression, chemotherapy, radiotherapy, immunotherapy, surgery, dehydration, and infection.

For every active cancer patient, the command center would continuously estimate:

• the risk of acute kidney injury in the next 24 to 72 hours
• the risk of severe AKI after cisplatin or another kidney-toxic therapy
• the probability that a rise in creatinine is immune checkpoint inhibitor nephritis rather than some other cause
• the probability of obstruction
• the risk of tumor lysis
• the risk of CKD progression after cancer therapy
• the risk of needing dialysis during a hospitalization
• the renal survivorship risk after cure or remission

The point that matters most to a clinician is the next one. The system would not merely alert. It would suggest the next useful action. Depending on the patient, that might be: repeat the urinalysis, check the urine microscopy, stop the PPI or the NSAID, adjust antimicrobial dosing, order a renal ultrasound, modify the cisplatin hydration, consult nephrology, consider a biopsy, or sit down and discuss modifying the cancer therapy. An alert tells you something happened. This tells you what to do about it.

This is the strongest caution, so it goes first. We can already predict acute kidney injury earlier than a physician might notice it. That has been shown. The trouble is what happens next. When you actually test the alerts in a controlled way, the patients often do no better.

A randomized trial published in BMJ tested electronic AKI alerts head to head against usual care. The primary outcome, a composite of AKI progression, dialysis, or death, occurred in 21.3% of the alert group and 20.9% of the usual-care group. That difference is not meaningful. The alert did not reduce harm. It is worth sitting with that: the technology worked, the prediction fired, and the patients were no better off.

A 2024 trial in JAMA went a step further. It tested a "kidney action team" that delivered early, individualized recommendations to the treating clinicians. The intervention succeeded at getting recommendations delivered and acted on more often. But it still did not significantly reduce the composite outcome of worsening AKI stage, dialysis, or mortality. Better delivery of advice, no measurable change in what happened to patients.

The lesson is the one any chief already knows from a career of running a service. A signal only helps if it is specific, if it is trusted, if it leads to a clear action, and if someone owns that action. A prediction with no owner and no plan is just more noise on a busy floor.

"Distribution shift" is a technical phrase, so here it is in plain English. A model learns from the patients it was trained on. If the world then changes, the model can quietly go stale without anyone noticing, because it is still confidently giving answers based on a world that no longer exists.

Cancer care changes faster than almost any field in medicine. New immunotherapies, antibody-drug conjugates, CAR-T products, bispecifics, targeted therapies, supportive-care protocols, and trial regimens arrive constantly. A model trained on last year's therapies may not understand this year's. And a model built at one cancer center may fail at another, because the patient mix, the formulary, the lab timing, the imaging frequency, and the admission practices are all different. A tool that is excellent in one place can be unreliable across the street. This is not a reason to avoid the tools. It is a reason to keep checking them against reality, the same way you would re-credential a clinician.

A nephrologist does not just want a score. He wants the why. Is the model worried because of the creatinine slope, the albumin, the cisplatin dose, an NSAID, the urine protein, hydronephrosis, hypotension, or a vancomycin level? Without the why, the number is unusable, because you cannot act on a cause you have not been shown.

Some modern models can show their reasoning. There is a technique, which you will see referred to as SHAP, that does one simple thing: for each prediction, it shows which findings pushed the risk up and which pushed it down, for that specific patient. So instead of "high risk," you get "high risk, driven mostly by the cisplatin dose and the falling magnesium." That is a model a clinician can argue with, and a model you can argue with is one you can trust. The explainable checkpoint inhibitor AKI work is built exactly this way.

AI is becoming a second layer of clinical attention. It is good at watching many variables at once, noticing risk earlier, quantifying images and slides, summarizing messy records, and standardizing the repetitive decisions. It is not yet good enough to replace judgment, to explain the tradeoffs, or to own the consequences of treatment. Those still belong to the physician, and the cautions chapter is the reason why.

In onco-nephrology the opportunity is unusually strong, because kidney injury in cancer is exactly the kind of problem AI suits: high stakes, data-rich, and causally tangled. The killer application is almost certainly not a generic AI doctor. It is a cancer-center renal intelligence system that tracks AKI risk, explains the likely causes, anticipates nephrotoxicity, helps triage which patients need a biopsy, supports the therapy decisions, and follows survivors for CKD long after the cancer is treated. Every piece of that is something this report has shown is either already real or within reach. The work is putting them in one place, connected to action, under a clinician's hand.