- Check the label before the model. Several of the strongest abstracts showed that the reference standard, not the algorithm, limits what a validation study can prove.
- Test models for consistency and calibrated uncertainty, not only accuracy. The same prompt gave different answers at deterministic settings, and chatbots quoted precise percentages the evidence cannot support.
- Clinical notes are becoming structured data. LLMs pulled symptoms from 455,000 hemodialysis notes with 98% sensitivity. That is the variable structured data has always missed.
- Delivery, not prediction, limits benefit. Two well-run cluster RCTs of AI and EHR alerts were negative. Risk scores paired with clinician feedback worked.
- Individual treatment benefit is real and hard to see. Causal models found patients who gain nothing from early dialysis, with a C-for-benefit of only 0.55.
Firecrawl drove the ASN abstract search: the full category 300 (Artificial Intelligence, Digital Health, and Data Science, 120 abstracts) plus 23 keyword searches across every other category. That produced 681 hits, 384 were read in full, and 157 were AI or digital health work. From those, 65 were shortlisted and Adam rated 53 on a 1 to 5 scale for level of interest.
Interest ratings (n = 53)
Bimodal: 24 rated 4 or 5, 22 rated 1 or 2, only 7 in the middle.
Mean interest by shortlist group
Scale 0 to 5. Group sizes: epistemics 14, deployment 10, research tools 11, novel 10, patient facing 6, dialysis ops 2.
TOP TIERThe reference standard is the weak link
You cannot validate a model more precisely than the definition you validate it against. These five abstracts each found the target itself was soft.
Development of a Large Language Model-Based Tool for Scoring Nephrology Clinical Reasoning
GPT-OSS 20B with a rubric prompt graded fellows' notes on a simulated AKI consult against a seven grader consensus (59 notes for tuning, 17 for testing). Accuracy was 87% for problem representation, 81% for differential diagnosis and 74% for diagnostic plan, where kappa ranged as low as 0.04. Where the human graders agreed least with each other, the model agreed least with them.
Takeaway: poor model agreement on complex reasoning may partly reflect an unreliable gold standard. Report inter-rater reliability next to every model metric.
Decoupling of Hemodynamic Instability and Patient-Reported Symptoms: A Deep Learning Analysis of 160,000 Hemodialysis Sessions
160,640 sessions in 418 patients. LSTM and CNN models predicted hemodynamic IDH well (ROC AUC 0.92, PR AUC 0.81) but symptoms poorly (ROC AUC 0.77, PR AUC 0.18). BP drops tracked fluid dynamics; symptoms tracked baseline frailty. 84.5% of sessions with SBP under 90 were asymptomatic, 90.3% of sessions with a drop over 30 mmHg were asymptomatic, and 73.4% of reported distress never met IDH criteria.
Takeaway: the composite IDH definition merges two different phenomena. Quality metrics and prediction models built on it inherit the confusion.
Diagnostic Aggregation and Semantic Drift in International Classification of Diseases (ICD)-Derived Kidney Phenotypes: Implications for Electronic Health Record (EHR)-Based and Genomic Research
585 kidney ICD codes from a 571,435 patient health system were mapped to Phecodes, SNOMED CT, HPO and OMIM, then rated by three blinded nephrologists. 39% of Phecode mappings were poor; the others had none. More codes collapsing into one concept meant lower fidelity (OR 0.33 per log increase). Glomerulopathy and CKD were the worst domains.
Takeaway: any EHR or genomic model trained on ICD derived kidney labels is learning a blurred target, especially in glomerular disease.
Electronic Health Record (EHR)-Based Phenotyping for Kidney Failure: A Composite Algorithm Outperforms Published Rules
Eight published EHR rules for identifying kidney failure were tested against an adjudicated cohort (2,590 with kidney failure, 5,407 without).
| Rule | Sens % | Spec % | PPV % |
|---|---|---|---|
| 1+ CKD 5 or ESRD diagnosis code | 44.2 | 94.6 | 79.6 |
| 1+ eGFR under 15 | 21.9 | 97.2 | 78.8 |
| 1+ transplant code | 17.0 | 98.3 | 82.7 |
| 1+ dialysis diagnosis or procedure code | 5.3 | 99.7 | 88.4 |
| 3+ dialysis codes over 30 days | 0.9 | 99.8 | 72.7 |
| Composite (any rule positive) | 53.9 | 91.9 | 76.1 |
Takeaway: even the composite misses nearly half of adjudicated kidney failure. How a study defines its cohort can decide its answer.
Validation of Large Language Model-Based Phenotyping for Kidney Stone Disease in the Mayo Clinic Biobank
Kidney stone phenotyping in the Mayo Clinic Biobank: LLM review of notes had PPV 88% versus 81% for ICD and survey data. Of chart-confirmed cases, 40 were found by both methods, 33 by the LLM only and 14 by ICD only. Combined NPV was 98.8%.
Takeaway: LLM phenotyping is a second instrument with a different error profile. Use both and study where they disagree.
TOP TIERHow models reason compared with clinicians
The question that matters is less "is the model right" than "does it express uncertainty where clinicians feel it, and does it give the same answer twice."
Artificial Intelligence (AI) Chatbots and Nephrologists' Reasoning Across Standardized Nephrology Case Vignettes
A resident, an early career and a senior nephrologist answered five glomerular cases alongside ChatGPT, Gemini, Claude, DeepSeek, Grok and Perplexity. Everyone got 4 of 5 diagnoses and everyone missed MCD in an elderly patient (though all listed it in the differential). Therapy agreement was high. Physicians gave prognosis as low, medium or high; the chatbots gave percentages, and they disagreed widely: FSGS remission 10 to 20% (Perplexity) versus 70 to 80% (Gemini); RRT in ANCA vasculitis 30 to 40% (ChatGPT) versus 70 to 85% (Claude).
Takeaway: spurious precision is a style. A percentage range from a chatbot is not a calibrated estimate.
Evaluating Large Language Models for CKD Medication Review: A Trap-Embedded Benchmark
50 synthetic CKD cases with 8 to 12 medications and ten embedded trap types, reviewed in triplicate at temperature 0 by GPT-5.4, Claude Sonnet 4.6, Gemini 3.1 Pro and Grok 4.1 Fast (600 outputs). F1 ranged 0.63 to 0.72. NSAIDs, metformin in advanced CKD and tacrolimus with azoles were caught 100% of the time. Binder timing was caught 14% of the time and statin gaps 65%. Safety critical error rates were 16 to 42% and hallucination rates 0 to 76%.
Takeaway: generic benchmarks test the famous traps. Specialty benchmarks find the quiet ones (timing, prevention), and reproducibility should be a reported safety metric.
Clinical Questions and Uncertainty in the ASN Open Forum: An Artificial Intelligence-Assisted Analysis
GPT-5.5 labeled 532 ASN Open Forum thread starters by disease area, question type and source of uncertainty. Glomerular disease and vasculitis led, then AKI and nephrotic syndrome. Most questions were about choosing treatment, and risk-benefit tradeoffs were the leading source of uncertainty. Human review found labels fully correct in 88% of a 50 post sample.
Takeaway: a working map of where nephrologists actually feel uncertain. That is where decision support should focus, and where AI overconfidence would do the most harm.
TOP TIERClinical notes are becoming structured data
Symptoms and patient experience live in free text. Cheap LLMs now extract them at scale with high accuracy, which also offers a way around the ICD problem above.
Artificial Intelligence (AI)-Based Extraction and Temporal Classification of Symptoms and Signs from Hemodialysis Notes
Nurse, technician and dietitian notes for 5,000 Fresenius patients in 2024 went through two GPT-5-mini passes: extract, then classify into 16 clinical and 14 timing categories. That yielded 192,261 symptoms or signs across 127,507 notes. On review of 200 notes: 96.6% correct classification, 98.0% extraction sensitivity, 98.8% timing accuracy.
Funding: Fresenius Medical Care.
Takeaway: a large dialysis organization is building structured symptom data from routine nursing notes. Paired with FR-PO1027, symptoms become a measurable outcome in their own right.
Context-Aware Classification of Fatigue Status in Electronic Health Record Notes Using Large Language Models in CKD and CVD Cohorts
Claude Sonnet 3.5 with few-shot prompting classified fatigue as affirmed, negated or not applicable in snippets from VA CKD and CVD notes. Physician agreement was kappa 0.85; the model reached 93.3% accuracy (affirmed F1 0.94, negated F1 0.96).
Takeaway: negation handling, the classic NLP failure, is now largely solved for common symptoms.
Large Language Model-Assisted Screening in a Systematic Review on the Use, Administration, and Interpretation of Patient-Reported Outcome Measures in Dialysis Randomized Controlled Trials
Eleven open-weight models and one frontier model screened 688 records for dialysis RCTs reporting PROMs. GPT-OSS-120b reached sensitivity 92.5%, specificity 94.0% and NPV 97.4%, matching Gemini 3 Flash. The 6.4% of records where its runs disagreed went to humans.
Takeaway: local open-weight models are good enough for this kind of work, and model self-disagreement is a usable routing signal for human review.
TOP TIERDelivery, not prediction, limits benefit
Two rigorous cluster RCTs of alerts were negative. The positive results came from feedback loops and from work that respected clinician attention.
Multihospital Electronic Decision Support for Drug-Associated AKI (MEnD-AKI): A Prospective Cluster-Randomized Clinical Trial
38 service clusters in 8 hospitals; 5,939 patients. A deep learning CDS triggered a pharmacist intervention for drug-associated AKI. MAKE30 was 50.8% with the intervention versus 42.9% with usual care (adjusted OR 1.05, 0.82 to 1.35). No secondary outcome differed. 52.3% of alerts were not actionable, and only 12% of urgent recommendations went by phone, the intended route; most went by email.
Funding: NIDDK.
Takeaway: the model fired; the system around it did not deliver. Actionability and real-time delivery should be designed and measured before the trial.
Clinical Decision Support Triggers to Improve CKD Screening and Detection in Primary Care: The ALMA-CKD Pragmatic Cluster Randomized Trial
65 Stockholm primary care centers randomized to EHR-embedded CKD triggers or standard care. Low physician engagement with the triggers.
| Outcome | Standard | CDS triggers | RR (95% CI) |
|---|---|---|---|
| Creatinine screening | 64.9% | 61.8% | 1.02 (0.88–1.18) |
| Albuminuria screening | 46.2% | 47.6% | 1.07 (0.94–1.23) |
| Creatinine retest after abnormal | 53.1% | 52.5% | 0.98 (0.89–1.08) |
| Albuminuria retest after abnormal | 37.6% | 37.5% | 1.01 (0.91–1.11) |
Funding: Stockholm healthcare system; unrestricted grant from AstraZeneca and Boehringer Ingelheim.
Takeaway: screening rates varied threefold across centers, and the triggers changed none of it.
Applying Risk Stratification to Improve Guideline-Directed Medical Therapy Uptake and Albuminuria Testing in a Large US Nephrology Practice
More than 100 providers and 50,000 patients. Klinrisk flagged patients with 25% or higher five-year progression risk, and providers got quarterly reports benchmarked to the practice. Among high-risk patients: SGLT2i 30.7% to 37.9%, GLP-1 RA 14.5% to 18.2%, MRA 13.5% to 15.9%, RAASi 61.2% to 63.5%. UACR testing rose from 43.0% to 50.7%. Gains were largest in high-risk patients.
Funding: Klinrisk Inc, AstraZeneca. Before and after design, not randomized.
Takeaway: feedback benchmarked against peers moved practice where pop-up alerts did not.
Large Language Models for CKD Guideline-Directed Medical Therapy
A secure GPT-5.4 workflow read structured and narrative EHR data ahead of visits. Of 20,470 patients screened, 2,996 were eligible and 945 (31.5%) got at least one recommendation: SGLT2i 57.5%, ACEi or ARB 28.7%, nonsteroidal MRA 15.0%, GLP-1 RA 9.8%. On review, 200 of 200 were appropriate and 199 of 200 safe.
Takeaway: excellent validation numbers, but no data yet on how many recommendations clinicians acted on. That is the question the two RCTs above say matters.
Real-Time Cumulative Sum-GFR Trajectory Monitoring Affords a 15-Month Predictive Lead Time for ESKD
A CUSUM control chart on outpatient eGFR (at least 9 values per patient), tuned on purpose to limit alert fatigue: sensitivity 96.4%, specificity 94.1%. 74.4% of patients who progressed were flagged while eGFR was above 45, with a mean 15.6 months warning before ESKD. The 10-year false positive rate was 16.8%.
Takeaway: classic statistical process control, interpretable and cheap, designed around clinician attention from the start.
Multimodal Deep Learning AKI Model Identifies Patients at Risk for Severe AKI and Adverse 90-Day Outcomes
A rare prospective validation: 148 inpatients flagged by the ESTOP 2.0 multimodal deep learning score at 2 centers. 38% developed severe AKI, 19% died in hospital and 63% were readmitted within 90 days.
Takeaway: the score finds sick patients. Whether finding them earlier changes outcomes is the untested next step.
TOP TIERIndividual benefit, and when not to treat
Causal machine learning found a group of patients who gain nothing from early dialysis. The signal is real, small, and depends on calibration more than discrimination.
A Multimodel Causal Survival Analysis Framework for Individualized Dialysis Decision Support in Patients with Advanced CKD: A Retrospective Cohort Study
PRISM combines four models to estimate five-year mortality benefit from starting dialysis within 90 days at eGFR 10 or below (n = 20,184, with spatial, temporal and external test cohorts). Patients were sorted into four zones by risk and predicted benefit. Early dialysis showed no survival advantage in the low-risk/low-benefit and high-risk/low-benefit zones; benefit was strongest in one zone (HR 0.28 to 0.38).
Takeaway: a candidate high-risk group for conservative kidney management, identified from routine data. Prospective testing is the obvious next step.
Comparison of Causal Machine Learning (ML) Models on Predicting Survival Benefit with Dialysis in Patients with ESKD: A Retrospective Cohort Study
Eight causal ML methods were compared on the same cohort with external validation (n = 6,040). Selected results on the external test set:
| Method | C-for-benefit | Calibration slope | Benefit recovered at 30% |
|---|---|---|---|
| Stacked ensemble | 0.554 | 0.77 | 75% |
| IPTW S-learner | 0.545 | 1.03 | 69% |
| T-learner | 0.547 | 0.47 | 64% |
| IPCW DML pseudo-outcome | 0.506 | 0.23 | 47% |
Takeaway: discrimination barely separates the methods. Calibration and transportability do, which is the authors' own conclusion.
Effectiveness of the Comprehensive Kidney Care Contracting (CKCC) Model on Kidney Treatment Outcomes and Medicare Expenditures
Medicare fee-for-service beneficiaries in CKCC practices versus geographically matched comparators, 2022 to 2023. CKCC practices improved permanent vascular access at HD start faster (+1.44 percentage points per quarter), and preemptive transplant rose only in CKCC-aligned CKD patients. Costs were $3,876 per year lower for CKD 4 to 5.
Takeaway: supportive of value-based kidney care. Check the methods before quoting, since effect sizes (Cohen d 2.9 to 5.8) are unusually large for an observational comparison.
Also rated highly: reaching patients between visits
A smaller cluster on low-cost channels for patient communication and education.
- FR-PO00475Using Artificial Intelligence (AI)-Assisted Short Message Service (SMS) Text Messaging to Support Patients with CKD: Early Lessons from a Community Nephrology Pilot30 CKD stage 2 to 4 patients, 6 months of AI-interpreted SMS check-ins. 83% responded, 700 responses including 347 readings and 74 clinical questions; NPS 67. Funded by Generated Health.
- FR-PO11204Patient-Facing Information on US Kidney Transplant Center Websites Exceeds Recommended Reading Levels: A National Analysis92,461 pages from 196 US transplant centers. Median reading grade 12.9 to 15.8 depending on the measure; 61 to 86% of pages above grade 12. A clear target for plain language rewriting.
- FR-PO00534Artificial Intelligence (AI)-Assisted Generation and Multilingual Translation of Nephrology Patient Education MaterialsGPT-5.3 turned 25 clinic notes into templated patient instructions and translated them into Spanish, Arabic, German and Hungarian. No critical errors; Hungarian weakest.
- SA-PO09754Artificial Intelligence (AI)-Assisted Educational Videos for CKRT and Intermittent Hemodialysis (iHD) Improve Trainee Knowledge and ConfidenceTen NotebookLM videos on CRRT and iHD for trainees. Correct CRRT ordering rose from 47% to 100% and understanding of clearance from 26% to 91%. Concepts were also reinforced with AI-generated songs.
The middle: rated 3
Seven abstracts were interesting but did not carry a theme.
- FR-PO00283Integrating Clinical Context to Improve Large Language Model Interpretation of Image-Based Nephropathology QuestionsLLM nephropathology accuracy rose from 43–62% with the image alone to 86–95% with the full vignette. The models read the story more than the slide.
- FR-PO00363Comparative Performance of Five Large Language Models in Systematic Review Abstract ScreeningFive OpenAI models screening review abstracts: newer was not consistently better, and every model missed relevant studies.
- SA-PO00853Effect of Prediction Horizons on Machine Learning Mortality Models in Patients on HemodialysisShortening a dialysis mortality model's horizon raised AUC but cut PR AUC from 0.30 to 0.20.
- FR-PO00493Patient Portal Messaging Among CKD Patients28.3 million MyChart messages from 2.3 million CKD patients in Epic Cosmos; volume fell with age and social vulnerability.
- FR-PO00323Comparing Large Language Model and Human Coding for Qualitative Analysis of Mindfulness in CKD: Insights from "The Mindful Kidney" ProjectClaude Code inductive coding of CKD interviews matched clinician coding (n = 5).
- FR-PO00253Tree-of-Thought and Graph-of-Thought Structure Improves Large Language Model Comprehension of Nephrology Graphical AbstractsGraphical abstracts with tree or graph structure were easier for a vision LLM to read.
- FR-PO00263Evaluating the Accuracy and Reasoning of Large Language Models (LLM) in Selecting Statistical Methodologies for Nephrology Clinical TrialsThree LLMs picked the right statistical method for six landmark trials. A ceiling result on a small set.
What rated low, and the pattern behind it
22 abstracts were rated 1 or 2. Many are good science. They score low because they are about a new device, image or molecule, or about research plumbing, rather than about how we know what we know or how care changes.
New sensors and opportunistic imaging
Clever capture methods without an epistemic or delivery question attached.
- SA-PO0077 2 Contact-free AV access monitoring (audio matched stethoscope, cosine 0.90)
- FR-OR047 2 DialySense contactless PD peritonitis sensor (96% sensitivity)
- SA-PO0072 2 Holographic WBC counting in pediatric PD effluent
- SA-PO0047 2 Zero-shot Claude Vision for donor biopsy adequacy (AUC 0.91 vs 0.80)
- SA-PO0062 1 Aortic calcification from lumbar X-rays
- FR-OR045 1 MRI kidney volume and hidden CKD risk
Bench science and drug discovery
AI used to generate hypotheses far from clinical decisions.
- SA-PO0092 1 Multi-agent AI co-scientist nominates diosmin
- TH-PO0174 1 LLM ranks ADPKD drug synergies
ICU and CRRT prediction
Acute care models outside dialysis operations.
Research plumbing
Useful workflow gains, but not about clinical reasoning.
Tools, QC and one-off cases
Includes two case reports filed under diagnostic reasoning. Individual bias anecdotes rated lower than system-level measurement problems.
- SA-PO0354 1 AKI wrongly blamed on contrast (anchoring)
- SA-PO0600 1 Lipemia producing a false metabolic acidosis
- FR-PO0031 1 OpenEvidence vs ChatGPT-5 for prior auth letters
- SA-PO0060 1 AI quality control for AI kidney volume measurements
- FR-PO0022 2 EHR AKI visualization tool adoption
- SA-PO1001 1 NEPHRON-AI tutoring chatbot usage
- SA-PO0080 1 Dysnatremia surveillance during a saline shortage
Direct resources
- Kidney Week 2026 abstract searchBrowse by category, date, author or keyword. Category 300 is Artificial Intelligence, Digital Health, and Data Science.
- Session: AI, Clinical Prediction, Decision Support, and Digital Kidney CareFriday poster session holding most of the LLM and CDS abstracts above.
- Session: AI, Kidney Imaging, Pathology, and Precision PhenotypingSaturday posters, including the Columbia phenotyping and Mayo biobank work.
- Session: AI in Kidney Care, From Pathology to PredictionThursday oral session, including the Brigham GDMT and trial prescreening talks.
- Session: Late-Breaking Research PostersMEnD-AKI and ALMA-CKD.
- ALMA-CKD trial registration (NCT06386172)Protocol for the Stockholm CDS trigger trial.
- Kidney Week 2026 plenary webcastsASN's recorded plenary lectures.
The models are getting good. The targets we grade them against are not. Before we deploy the next prediction tool, we should audit the label, test the model twice, and design the delivery path as carefully as the algorithm.