Cheungpasitporn, Wisit

Introduction: Large language models (LLMs) match or exceed physician accuracy on benchmark diagnostic tasks. Whether higher accuracy reflects human-like diagnostic behavior, which bears on how safely clinicians can supervise these systems, is rarely assessed. Objectives: To determine whether diagnostic accuracy and alignment with human reader responses move together across widely accessible LLMs under default, point-of-care conditions. Methods: This cross-sectional evaluation submitted the 29 most recent New England Journal of Medicine (NEJM) case challenges to 5 publicly accessible LLMs (GPT-5, Claude, Google Gemini, DeepSeek, and Grok) through each model’s web interface (November 5-8, 2025) at default settings, one query per case, no system prompt. Each case offered a clinical vignette and 6 candidate diagnoses; published human reader response distributions (4386-17 714 per case; median, 7732) were the comparator. Diagnostic accuracy against the published gold-standard diagnosis, compared with the human reader majority by the exact McNemar test; agreement by Cohen κ; alignment by a human-error similarity index (the share of model errors on cases the reader majority also missed) and by the Jensen-Shannon divergence (JSD) between each model’s response and the reader distribution; and ordinary least squares regression of per-case JSD on normalized reader-disagreement entropy. Results: Reader majority accuracy was 48.3% (95% CI, 31.4-65.6). All 5 reached or exceeded this value, from 51.7% for DeepSeek (95% CI, 34.4-68.6) to 82.8% for Grok (95% CI, 65.5-92.4); Grok (P = .02), GPT-5 (P = .004), and Gemini (P = .04) significantly exceeded the reader majority, whereas Claude and DeepSeek did not. Grok, the most accurate model, showed the weakest alignment (similarity index, 0.40; mean JSD, 0.50 bits), with 3 of 5 errors on cases readers answered correctly. GPT-5 showed the strongest alignment (index, 1.000), failing only on cases that also challenged readers. For 4 of 5 models, divergence rose with reader-disagreement entropy; DeepSeek was the exception. Conclusions and Relevance: In this evaluation, diagnostic accuracy and alignment with human reader responses diverged across LLMs. Because a model can outperform clinicians in aggregate yet fail where oversight is least likely to detect it, evaluation of clinical AI should weigh error structure alongside accuracy.

Background Public awareness and regional engagement with cardiorenal syndrome (CRS) remain poorly understood. We examined temporal trends and geographic patterns of U.S. public interest in CRS and their relationship with disease burden. Methods Google Trends™ data (2004–2025) annual relative search interest (RSI) for CRS. CRS-related mortality was derived from CDC WONDER (2007–2020) using hypertensive heart and chronic kidney disease deaths. Temporal trends were assessed via augmented Dickey-Fuller tests and linear regression. Associations between RSI and mortality were evaluated using Pearson correlation, AR(1) modeling, first-difference, and cumulative change analysis. State-level correlations were examined with and without adjustment for income, education, and insurance coverage. Results Both RSI and mortality exhibited strong non-stationary upward trend ( p = 0.98 and 0.42). Raw data showed a strong correlation analysis (r = 0.79, p = 0.001); but first-differenced year-over-year changes were not significant (r = −0.36, p = 0.23), consistent with AR(1) modeling ( β =1.58, p = 0.11). Cumulative changes remained strongly correlated (r = 0.80, p = 0.001), indicating parallel long-term trajectories. Findings were consistent with varying ICD-10 code definitions for CRS-related mortality. Geographically, RSI was highest in West Virginia and lowest in Texas and Georgia. State-level RSI showed no correlation with age-adjusted mortality, both unadjusted (r = −0.09, p = 0.52) and after adjustment for median income, educational, and insurance coverage ( β =0.03, p = 0.75). Conclusion Public interest in CRS has grown in parallel with mortality, reflecting its increasing epidemiological relevance. However, the lack of geographic correlation between search interest and mortality—even after socioeconomic adjustment—suggests awareness gaps and highlights opportunities for targeted education and resource allocation.

Artificial intelligence (AI)–based protein structure prediction has rapidly entered nephrology, providing plausible atomic models for channels, transporters, receptors and scaffolds that previously lacked structural information. AlphaFold and RoseTTAFold now underpin mechanistic interpretations of inherited kidney diseases, including autosomal-dominant polycystic kidney disease, cystinosis and Alport syndrome, and support experimental design in cryo-electron microscopy and cryo-electron tomography. For nephrologists, this means that renal proteins no longer need to be treated as black boxes: genetic, biochemical and physiological data can now be mapped onto concrete three-dimensional scaffolds. However, the apparent precision of AI models can be misleading. Renal proteins frequently fall within underrepresented classes, including multipass membrane proteins, intrinsically disordered scaffolds, and human-specific isoforms with shallow multiple sequence alignments, where confidence scores may not accurately reflect model reliability. Current predictors typically deliver static snapshots, whereas transporters, channels and slit-diaphragm complexes function as dynamic ensembles embedded in specific lipid and macromolecular environments. Attempts to use AlphaFold directly for variant classification or to estimate changes in thermodynamic stability or function show only modest agreement with mutagenesis data, particularly for gain-of-function and dominant-negative alleles that depend on subtle conformational and interaction changes. Common pitfalls include overinterpreting oligomeric stoichiometry inferred from AlphaFold Multimer, reading too much into domain orientations with high predicted aligned error, and dismissing low-confidence regions as modelling failures rather than recognising intrinsically disordered, potentially regulatory segments. Integrative approaches combining cryo-electron tomography, crosslinking mass spectrometry and molecular dynamics can mitigate these limitations but remain labour-intensive and low-throughput. An explainable, hypothesis-driven use of AI that treats predicted structures as working scaffolds to be cross-validated against imaging, biophysics and functional assays offers the most reliable path towards a genuinely mechanistic nephrology and responsible clinical translation of AI-based structural insights.

Background Intermittent fasting (IF) represents a dietary strategy with known metabolic benefits, however its safety and applicability in patients with chronic kidney disease (CKD) remains poorly understood. AI chatbots’ reliability for clinical and dietary guidance in clinically complex areas remains underexplored. No study has examined AI performance on IF in patients with CKD. This study evaluated the accuracy, consistency, readability, and quality of responses generated by ChatGPT-5 and Claude 4.5 Sonnet regarding IF in CKD. Methods Twenty-five patient-centered questions spanning seven clinical domains were developed by two board-certified nephrologists and submitted independently to ChatGPT-5 and Claude 4.5 Sonnet twice per model. Prompts were submitted under standard and unmodified user conditions without customized system instructions to represent a zero-shot evaluation that is illustrative of typical patient use. Responses were evaluated for accuracy, consistency (comparatively between responses), and appropriateness by the evaluators. Response readability and quality were assessed using the Flesch-Kincaid Grade Level (FKGL) formula and the DISCERN instrument respectively. Results Both ChatGPT-5 and Claude 4.5 Sonnet demonstrated identical accuracy scores during the first question prompts with notable improvement upon re-prompting after 24-hours. Both models demonstrated an identical internal consistency at 90%. ChatGPT-5 produced responses at the middle school level, better aligned with the US adult health literacy benchmark, while Claude 4.5 Sonnet generated responses at the high school level. DISCERN scoring was comparable between chatbots indicative of good response quality. Both chatbots had similar performance regarding individual subdomains. Conclusions Both chatbots illustrated clinically accurate and cautious responses to questions regarding IF in CKD. ChatGPT-5 responses were better suited to the average health literacy of the general United States adult population. Despite our findings, we do not recommend either chatbot be used as the primary source for clinical counseling regarding IF in CKD. It is important for clinicians to help to guide patients toward evidence-based clinical information.

Rare diseases affect over 300 million individuals worldwide yet remain underdiagnosed and poorly characterized due to fragmented data, small cohorts, and phenotypic heterogeneity. Advances in artificial intelligence (AI) are enabling integration of genomics, imaging, electronic health records, and patient-generated data to support diagnosis, phenotyping, prognosis, and therapeutic discovery. In kidney care, these capabilities are reflected in tools for genomic variant prioritization, AI-assisted histopathology, and integrated risk stratification models for rare and complex kidney diseases. This review synthesizes current AI applications across the rare disease continuum and proposes a clinically grounded framework to distinguish exploratory models from systems that are methodologically robust and operationally deployable. We highlight advances that address data sparsity and heterogeneity, alongside persistent challenges in validation, generalizability, equity, and workflow integration. Finally, we outline future directions, including federated learning, digital twins, and AI-driven clinical decision agents, as pathways toward precision-guided, workflow-integrated rare disease care.

Background Tacrolimus dosing after kidney transplantation is complex, highly individualized, and prone to variability, which can impact graft outcomes. While machine learning (ML) approaches have been used primarily to predict tacrolimus concentrations, large language models (LLMs) may enable protocol-constrained clinical decision support by generating dosing recommendations aligned with established treatment guidelines. Methods We developed TacroDose AI, a protocol-constrained LLM based on GPT-4, to generate tacrolimus dosing recommendations aligned with institutional guidelines. Using 300 structured simulated clinical scenarios, we evaluated protocol adherence, error patterns, and reproducibility of model-generated dosing recommendations. Following review of initial outputs, a refined prompt ( TacroAI 2.0 ) was implemented to strengthen protocol constraints and structured output verification, and performance was reassessed. Results TacroDose AI achieved 72.9% protocol adherence, with 27.1% protocol deviations, including 13.9% sentinel errors with potential clinical significance. Prompt refinement substantially improved performance: TacroAI 2.0 achieved 91.7% adherence, reduced sentinel errors to 1.0%, and improved reproducibility from 77.3 to 86.7% (all p < 0.01). Although rounding discrepancies increased, these represented minor numerical differences without clinical relevance. Performance improvements were observed across subtherapeutic, therapeutic, and supratherapeutic tacrolimus scenarios. Conclusion Protocol-constrained LLMs can produce guideline-concordant tacrolimus dosing recommendations with high adherence and improved reliability in simulated clinical settings. These findings highlight a potential framework for integrating generative AI with rule-based clinical protocols to support safe medication management. With further validation using real-world data, such systems could be integrated into electronic health records to support clinician-supervised, protocol-based immunosuppressive dosing in transplant care.

Background The clinical heterogeneity of systemic lupus erythematosus exceeds the resolution of conventional disease activity instruments. Artificial intelligence offers the analytical infrastructure to engage with this complexity, yet the translation of AI models into clinical practice remains limited. Methods This review critically appraises the current evidence for AI applications across the SLE clinical workflow, including computational phenotyping, diagnosis, disease activity monitoring, organ-specific predictive modelling, and treatment personalization. Studies were evaluated for external validation, prospective testing, algorithmic fairness, explainability, and regulatory status. Findings AI applications across the SLE clinical workflow show uneven methodological maturity. Diagnostic and monitoring tools include the SLERPI index, with multinational external validation, and an LSTM flare-prediction model (C-index 0.897). Treatment personalization is anchored by serum IgA2 anti-dsDNA, the only SLE response biomarker validated across independent trials, and by multi-stain deep learning on renal biopsies (AUC 0.84, three external centers). Predictive modelling for organ-specific manifestations has progressed unevenly: cardiovascular risk stratification (SLECRISK) and a multicenter thrombocytopenia model are the most mature, while neuropsychiatric, gastrointestinal, and ocular applications remain single-center proofs of concept. External validation is the exception across SLE prediction models, no AI tool has received regulatory clearance, and populations most affected by SLE remain underrepresented in training data. Conclusions AI demonstrates genuine analytical capability in SLE but the translation gap is defined by insufficient validation, limited explainability, and absent equity evidence. Closing it will require fewer published models and more validated tools, optimization for clinical impact rather than discrimination alone, and demonstration of performance equity across the demographic spectrum of the disease.