Citations

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.

OBJECTIVE The objective was to synthesize denominator-based evidence on indications, techniques, and postoperative outcomes of cervicomedullary decompression (CMD), spinal canal stenosis surgery, and cerebrospinal fluid (CSF) diversion for hydrocephalus in achondroplasia. METHODS This systematic review followed PRISMA 2020 and was prospectively registered (PROSPERO CRD42025626915). MEDLINE (via PubMed), Embase, Scopus, Web of Science, and the Cochrane Library were searched from inception through November 2025. The authors included observational studies of patients with achondroplasia who underwent CMD, spinal canal stenosis surgery, or CSF diversion for hydrocephalus with extractable outcomes and definable denominators. Two reviewers independently screened studies and extracted data. Risk of bias was assessed using ROBINS-I. Random-effects single-arm meta-analyses were performed for prespecified endpoints (any improvement, complete resolution, and mortality) only in cohorts with fully categorizable symptom status. RESULTS The authors included 38 observational studies (1442 patients): 18 CMD studies with 600 patients, 18 spinal stenosis studies with 686 patients, and 4 hydrocephalus studies with 156 patients. No study was at low risk of bias. For CMD (mean age 3.5 years; mean follow-up 27.8 months), any improvement was observed in 93% (95% CI 84%–97%) of patients and complete resolution in 44% (95% CI 16%–77%), and pooled mortality was 0% (95% CI 0%–5%). For spinal stenosis (mean age 36.6 years; mean follow-up 53.1 months), any improvement was observed in 92% (95% CI 78%–97%) and complete resolution in 30% (95% CI 16%–51%), and pooled mortality was 1% (95% CI 0%–2%). Hydrocephalus management was predominantly ventriculoperitoneal shunting (74.4%), with endoscopic third ventriculostomy in 23.1%, but symptom outcomes were rarely classifiable (9.6%), precluding strategy-level inference. CONCLUSIONS CMD and spinal stenosis surgery were frequently followed by symptom improvement in published series, whereas complete resolution was variably defined and inconsistently reported. Hydrocephalus studies described practice patterns but lacked outcome reporting adequate for comparing diversion strategies. Validated patient-reported outcomes were rarely reported, limiting inference on health-related quality of life.

Objective: The aim of this study was to evaluate outcomes of fenestrated-branched endovascular aortic repair (FB-EVAR) of thoracoabdominal aortic aneurysms (TAAAs) and complex abdominal aortic aneurysms (CAAAs) in patients actively smoking at the time of repair compared with previous smokers and non-smokers. Methods: Single-center retrospective review of elective FB-EVARs for TAAA and CAAAs between 2007 and 2023 was conducted. Demographics, comorbidities, aneurysm extent/ morphology, technical details, and peri-operative outcomes were collected. Patients were grouped into 1 of 3 groups: active smokers, prior smokers, and never smokers. Patients who stopped smoking within 6 weeks to surgery were considered active smokers. Rupture and symptomatic aneurysms were excluded. Endpoints included early (<30 days) mortality and major adverse events (MAEs). Results: There were 676 patients (73% male, mean age 74 ± 7.8). There were 141 (21%) active smokers, 455 (67%) prior smokers, and 80 (12%) never smokers. Active smokers were significantly younger than prior or never smokers (71 ± 6.7, 75 ± 6.9, and 76 ± 12, respectively; P < .001). Cardiovascular risk factors were similar, except active smokers were significantly more likely to have chronic obstructive pulmonary disease (60% vs 42% vs 6.2%, P < .001) and peripheral arterial disease (31% vs 24% vs 10%, P = .002) when compared to prior and never smokers, respectively. Patients who never smoked had significantly more extent II aneurysms (34%) than prior (18%) and active (15%) smokers, P = .001. Univariable and multivariable logistic regression analyses showed no statistically significant associations between smoking status and any specific peri-operative adverse events following surgery. Conclusion: Active smoking status does not appear to affect peri-operative outcomes after FB-EVAR for complex aortic aneurysms. Although smoking cessation must be strongly encouraged, repair should not be delayed due to active smoking status. Clinical Impact These data suggest that active smoking, in isolation, should not be considered a contraindication to elective FB-EVAR when anatomy is suitable and patients are otherswise optimized. This is particularly relevant for high-risk patients who may have limited alternative treatment options. Proceeding with FB-EVAR in active smokers does not appear to compromise early safety or procedural efficacy.

BACKGROUND Intraneural perineuriomas (INPs) are rare benign peripheral nerve sheath tumors with characteristic clinical and MRI features. Their 18 F-fluorodeoxyglucose (FDG) positron emission tomography (PET) appearance has not been characterized in a patient series. The authors evaluated FDG uptake patterns in INP. OBSERVATIONS The authors retrospectively reviewed that data of patients with confirmed INP who had MRI and FDG PET studies that included the affected nerve. Eight patients were included (5 male and 3 female; mean age 30.8 years, range 16–61 years). All lesions showed mildly increased FDG uptake relative to the contralateral nerve (lesion maximum standardized uptake value [SUVmax]: mean 1.5, median 1.2, range 1.1–3.5; contralateral nerve SUVmax: mean 0.9, median 0.9, range 0.8–1.0) and skeletal muscle background (SUVmean: mean 0.8, range 0.5–0.9). Six patients had denervation edema and/or fatty atrophy, without appreciably increased FDG uptake in the affected musculature. No additional abnormal hypermetabolic foci were identified. LESSONS INPs generally demonstrate low-level FDG uptake. PET is best regarded as a complementary test when already available or when metabolic characterization may refine the differential diagnosis; unexpectedly high uptake should prompt reconsideration of the diagnosis or additional evaluation. https://thejns.org/doi/10.3171/CASE26698

Abstract Background: Sexual and gender minority (SGM) representation is growing in the United States. This study estimated the future SGM cancer burden amidst shifting cancer patterns. Methods: Age-specific incidence data for breast, prostate, colorectal, lung, and melanoma were obtained from the Surveillance, Epidemiology, and End Results Program’s 12-registry database. Nordpred age-period-cohort models projected incidence through 2040 to 2044. Rates were applied to US Census population estimates and generation- and sex-specific SGM identification proportions from Gallup’s 2023 to 2024 surveys. Prophet was used for sensitivity analysis. Results: At ages 55 to 59, breast cancer cases among Millennial SGMs were 3,672% higher than among Silent Generation SGMs (34,850 vs. 924 cases), melanoma cases were 1,061% higher (5,526 vs. 476 cases), and colorectal cancer cases were 2,206% higher (23,594 vs. 1,023 cases). Compared with Generation X, Millennial prostate cancer cases were 142% higher (8,100 vs. 3,344 cases), and lung cancer cases were 169% higher (5,451 vs. 2,029 cases). By 2040 to 2044, Millennials were projected to account for 68.4% of breast cancer cases and 59.7% of colorectal cancer cases among SGMs. Conclusions: At ages 55 to 59, the 5-year cancer burden was projected to be 15 times higher among Millennials than Silent Generation SGMs (77,521 vs. 5,134 cases), driven primarily by the larger SGM population in younger generations. Oncology systems will need to prepare for the increase in cancer care demand. Impact: These findings identify a substantial generational shift in the SGM cancer burden and may inform oncology workforce capacity, culturally responsive cancer care, and healthcare system preparation for this growing population.

Background: Normothermic machine perfusion (NMP) may reduce preservation injury in donation-after-circulatory-death (DCD) liver transplantation, but its association with patient-centered recovery and longer-term graft survival remains uncertain. Study Design: We retrospectively analyzed 306 consecutive adult DCD liver transplant recipients at a single center from January 2019 through December 2022, comparing NMP (n=95) with static cold storage (SCS; n=211). The primary endpoint was days alive and out of hospital at 365 days (DAOH365). Quasi-Poisson regression evaluated factors associated with DAOH365, and Cox regression evaluated graft survival. Propensity-matched and calendar-year sensitivity analyses assessed measured confounding. Results: Median DAOH365 was 357.0 (352.0, 360.0) days with NMP versus 353.0 (336.0, 359.0) with SCS (p<0.001), but preservation strategy was not independently associated with DAOH365 (adjusted rate ratio 1.02; p=0.331). NMP recipients had less early allograft dysfunction (35.8% vs 59.7%; p<0.001), ischemic cholangiopathy (4.2% vs 23.7%; p<0.001), and retransplantation within 2 years (0.0% vs 6.6%; p=0.007). Two-year patient survival did not differ significantly. NMP was associated with improved twoF-year graft survival (HR 2.9, 95% CI 1.23–6.87; p=0.015), an association maintained in propensity-matched and calendar-year sensitivity analyses. Adjusted total 1-year costs did not differ (p=0.512). Conclusions: In this single-center DCD cohort, NMP was associated with superior 2-year graft survival but not independently with DAOH365. These hypothesis-generating findings support multicenter prospective evaluation of NMP using both graft survival and patient-centered recovery endpoints.

BACKGROUND: The right ventricular (RV) response to pressure overload is a crucial determinant of outcomes in pulmonary arterial (PA) hypertension. Sotatercept improves clinical outcomes, but its effects on the right ventricle have been inconsistent. We evaluated the effects of sotatercept on load-independent measures of RV contractility, RV–PA coupling, and right heart deformation at rest and exertion in patients with PA hypertension. METHODS: In this prospective trial among participants with PA hypertension recruited between July 2024 and May 2025, we performed simultaneous echocardiography and invasive pressure measurements during rest and exercise to evaluate changes in RV contractility (end-systolic elastance [Ees]), afterload (PA elastance [Ea]), RV–PA coupling (ratio of arterial Ea to Ees), and right heart deformation, including RV and right atrial strain, before and after 24 weeks of sotatercept using mixed models. RESULTS: Among 30 participants (n=21; 70% female; mean age, 49.3 years [SD 13.5]), sotatercept decreased PA Ea. These changes were associated with reduced RV contractility, PA pulsatility index, and the maximal rate of RV pressure increase over time (dP/dt max ; P <0.001 for all). The decrease in Ees was associated with a decrease in PA Ea (r=+0.70 [95% CI, +0.44 to +0.85]; P <0.0001) with preserved RV–PA coupling (mean difference, +0.08 [95% CI, −0.11 to +0.26]; P =0.42). Although systolic indices measured at the base of the right ventricle (tricuspid annular plane systolic excursion and systolic velocity of the lateral tricuspid annulus) decreased after sotatercept use, there was an increase in RV free wall–dependent systolic indices, including resting and exercise RV free wall strain (mean difference, +6.2% [95% CI, +4.5 to +7.8]; P <0.0001) and fractional area change ( P =0.0001). Beyond the right ventricle, there was also improved right atrial filling and contractile performance across rest and exercise ( P <0.0001 for all). CONCLUSIONS: Sotatercept decreases RV contractility in proportion to reduced afterload, while preserving the RV–PA coupling relationship. Despite decreased RV contractility, global measures of right heart performance, such as RV free wall strain and RA strain, improve, alongside redistribution of RV deformation from annular motion to free wall shortening. These data suggest that sotatercept causes an energetically favorable reduction in resting RV contractility to match the lower afterload, which does not appear physiologically detrimental when considering the concurrent improvement in RV free wall contractile mechanics and exertional right heart reserve. REGISTRATION: URL: https://www.clinicaltrials.gov ; Unique identifier: NCT06409026.

Background: Reliable access to postoperative guidance between clinic visits remains inconsistent, and the usefulness of large language models depends on how clinical knowledge is structured and retrieved. Knowledge graph-only (KG-only), vector retrieval-augmented generation (RAG-only), and hybrid graph-retrieval (GraphRAG) architectures have not been directly compared under controlled conditions for postoperative care. Objective: It was to compare these three pipelines on a curated postoperative question-answering benchmark using automated grounding metrics and blinded clinician review. Methods: A benchmark of 250 expert-authored questions (200 in-scope, 20 emergency-intent, 30 out-of-scope) was evaluated on an identical knowledge corpus using Gemini 2.5 Flash for generation and Gemini Embedding-001 for retrieval. Three blinded clinicians rated all 750 outputs; the primary outcome was majority-consensus accuracy across all 250 questions. GraphRAG is a multi-component pipeline that also includes query expansion, a second retrieval stage, evidence fusion, and a monotonic RAG fallback, and is therefore evaluated as a complete hybrid system rather than as graph retrieval in isolation. Results: Overall accuracy was 96.8% (95% CI: 93.8–98.4) for GraphRAG, 94.8% (95% CI: 91.3–96.9) for RAG-only, and 56.0% (95% CI: 49.8–62.0) for KG-only. Both retrieval-augmented systems outperformed KG-only (adjusted p < 0.001). The GraphRAG versus RAG-only difference was 2.0 percentage points, arose from five discordant question pairs and did not reach the Bonferroni-adjusted threshold (exact McNemar p = 0.062; adjusted p = 0.188). The monotonic RAG fallback determined 44.8% of GraphRAG outputs (112 of 250 queries). GraphRAG showed higher context recall (0.851 versus 0.796; adjusted p < 0.001), whereas RAG-only showed higher context precision. Estimated hallucination risk, derived from an LLM-as-a-judge procedure using Gemini 2.5 Flash rather than independent clinical adjudication, was 0.108, 0.132, and 0.436, respectively; the GraphRAG versus RAG-only difference was not significant (adjusted p = 0.771). Inter-rater agreement was high (Fleiss’ kappa = 0.854). Conclusions: Retrieval augmentation, rather than the choice of retrieval architecture, accounted for most of the difference observed. The incremental advantage of GraphRAG over a well-implemented RAG-only system was small, was not statistically significant after correction and reflects the complete hybrid system rather than graph augmentation alone. Emergency-intent accuracy was 75.0% for both retrieval-augmented systems; the architecture is therefore suited to supervised clinical workflows with clinician review, not autonomous patient-facing triage. Prospective validation on authentic patient messages and formal component ablations are required.