Citations

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.

Abstract The liver is organized into tightly regulated zones with distinct metabolic functions but zonation erodes with age. Cellular senescence contributes to aging and liver diseases, however, its impact on aging biology is ill-defined. As part of The Cellular Senescence Network Consortium, we used multiple spatial transcriptomics approaches (GeoMx, Visium, CosMx) with snRNA-seq to profile senescence signatures, zonation markers, and metabolic pathways in livers from wild-type (WT) mice of multiple ages. We observed a loss of canonical zone signatures in aged mouse livers characterized by “expansion” of midlobular (zone 2) marker gene expression, accompanied by diminished expression of zone 3 marker genes by middle-age (18 months), indicative of loss of cell identity. Multiple analytic approaches identified distinct age-, zone- and sex-specific senescence signatures, which were significantly associated with zonation markers changes. This was recapitulated in Ercc1 mutant models of accelerated senescence, supporting a causal role of senescent cells in liver aging. A “no-zone” hepatocyte-like cluster expanded with age and with the strongest Senescence-Associated Secretory Phenotype (SASP) profile. Gene expression profiles from senescent hepatocytes implicate decreased WNT signaling and increased BMP as contributing to age-related loss of zonation. Together, these data elucidate the role of senescent cells in driving aging biology in non-diseased liver through disruption of cell:cell signaling and the loss of metabolic and cell identity gene expression necessary for hepatocyte function.

Importance Elevated low-density lipoprotein cholesterol (LDL-C) is a modifiable risk factor for cardiovascular disease, the leading cause of premature death worldwide. Assessing the LDL-C–related burden is critical for guiding prevention and treatment strategies. Objectives To estimate the global, regional, and national burden of ischemic heart disease and ischemic stroke attributable to elevated LDL-C (relative to 35-54 mg/dL) from 1990 to 2023 and to quantify the contributions of population growth, aging, risk-deleted burden, and exposure changes to burden trends. Design, Setting, and Population This comparative risk assessment, part of the Global Burden of Disease Study 2023, estimated population-level LDL-C exposure and associated health loss in 204 countries and territories. Mean LDL-C levels were estimated using spatiotemporal gaussian process regression based on 806 studies across 161 countries. Relative risks were derived from meta-analyses of 38 randomized clinical trials. Population-attributable fractions for deaths and disability-adjusted life-years (DALYs) were estimated by age and sex for adults aged 25 years or older from 1990 to 2023, with 95% uncertainty intervals. Exposure Population-level LDL-C concentrations. Main Outcomes and Measures Population-attributable fractions, counts, and rates (all ages and age standardized per 100 000) of LDL-C–attributable deaths and DALYs from ischemic heart disease and ischemic stroke, with uncertainty intervals. Results In 2023, elevated LDL-C accounted for 3.6 million deaths (95% uncertainty interval, 2.2-5.4 million; 6.0% of global mortality) and 90.7 million DALYs (95% uncertainty interval, 58.9-123.3 million; 3.2% of DALYs). Although global all-ages rates remained stable, age-standardized death and DALY rates decreased by 45.6% and 39.5%, respectively, since 1990. In 2023, age-standardized LDL-C–attributable DALY rates were highest in Eastern Europe and lowest in high-income Asia-Pacific. One-third of the global LDL-C burden occurred in India and China. Population growth and aging drove the increasing burden, with notable regional disparities in LDL-C exposure and risk-deleted DALY rates shifting toward middle-sociodemographic settings. Conclusions and Relevance Despite declining age-standardized rates, the absolute LDL-C burden has increased since 1990 due to demographic changes and has shifted toward middle-sociodemographic countries. Measurement and surveillance gaps persist. Strengthened prevention, diagnosis, and treatment access strategies are essential to mitigate the health burden of LDL-C.

e18624 Background: Invasive fungal infections (IFIs) are a major cause of morbidity and mortality in patients with hematologic malignancies. Although meta-analyses identify systemic corticosteroid exposure as a strong relative risk factor for IFI, the absolute burden attributable to corticosteroids across heterogeneous baseline risk settings remains poorly defined. We applied meta-analysis-based projection modeling to translate relative risk estimates into absolute, population-level IFI outcomes. Methods: Using pooled odds ratios for IFI associated with systemic corticosteroid exposure from a published meta-analysis (OR 2.84; 95% CI 1.42–5.70), we projected IFI incidence under a proportional odds framework. The primary analysis assumed a baseline IFI incidence of 15% in patients not receiving corticosteroids, with sensitivity analyses conducted across baseline incidences ranging from 10% to 25%. IFI-related mortality was estimated using a case-fatality rate of 45%. Outcomes were expressed as projected IFI incidence, excess IFIs and IFI-attributable deaths per 1,000 patients, population attributable fractions, and one-way sensitivity analyses. Results: At a baseline IFI incidence of 15%, systemic corticosteroid exposure increased projected IFI incidence to approximately 34%, corresponding to an excess of ~190 IFIs per 1,000 patients. This translated into approximately 80 excess IFI-attributable deaths per 1,000 patients, with a plausible range of 25–160 based on uncertainty in the pooled effect estimate. Sensitivity analyses demonstrated a monotonic increase in projected IFI incidence with rising baseline risk, with corticosteroid-associated IFI incidence approaching 50% at higher baseline levels. Population-attributable fraction modeling indicated that when corticosteroid use prevalence exceeded 40%, nearly half of IFIs in the population were attributable to corticosteroid exposure. One-way sensitivity analyses identified the corticosteroid effect size as the dominant driver of excess IFI mortality, exceeding the influence of baseline incidence or IFI case-fatality assumptions. Conclusions: Projection modeling based on meta-analytic risk estimates demonstrates that systemic corticosteroid exposure is associated with a substantial and clinically meaningful absolute increase in IFIs and IFI-related mortality. Translating relative risk estimates into absolute population-level outcomes highlights the potential magnitude of preventable infectious harm associated with corticosteroid use and provides actionable insight to inform risk-mitigation strategies in high-risk hematologic populations.

e18602 Background: Chimeric antigen receptor T-cell (CAR-T) therapy has transformed the treatment of hematologic malignancies but is frequently complicated by cytokine release syndrome (CRS) and immune-mediated neurotoxicity (NS). As CAR-T programs expand, population-level projections of treatment-related toxicities are essential for program planning and risk mitigation, yet remain poorly characterized. Methods: We performed a cohort-based forecasting analysis using pooled incidence rates of CRS and NS derived from published meta-analyses of CAR-T clinical trials. Expected numbers of toxicity events were projected for hypothetical cohorts of 100, 500, and 1,000 CAR-T–treated patients. Uncertainty was quantified using binomial 95% confidence intervals. Disease-specific projections were generated by applying stratified pooled toxicity rates for hematologic malignancies and solid tumors. Results: For a projected cohort of 1,000 CAR-T–treated patients, forecasts estimated approximately 770 patients would experience CRS of any grade, 290 would develop grade ≥3 CRS, 400 would experience NS of any grade, and 280 would develop grade ≥3 NS. Toxicity counts increased proportionally with cohort size and demonstrated predictable uncertainty bounds. When stratified by disease type, hematologic malignancies exhibited substantially higher toxicity burdens than solid tumors across all endpoints. At a cohort size of 1,000 patients, hematologic malignancies were associated with markedly higher rates of both CRS and severe neurotoxicity, whereas solid tumors demonstrated consistently lower toxicity burdens. Conclusions: Using pooled clinical trial data, this forecasting model provides quantitative estimates of CAR-T–associated toxicity burdens at the population level. Severe CRS and neurotoxicity remain common and scale predictably with program size, particularly in hematologic malignancies. These findings offer a pragmatic framework for anticipating resource utilization, toxicity management needs, and safety planning as CAR-T therapy continues to expand.

e18536 Background: Nucleophosmin 1 (NPM1) mutation is traditionally associated with a favorable prognosis in acute myeloid leukemia (AML), leading to classification of these patients within the European LeukemiaNet (ELN) favorable risk group. However, emerging evidence indicates that co-occurring secondary-type mutations (STMs), including ASXL1 , RUNX1 , EZH2 , and SRSF2 , may substantially modify this prognosis. While meta-analyses have quantified the relative risk associated with STMs using hazard ratios, the corresponding absolute survival impact remains unclear, limiting clinical interpretability and decision-making. Methods: We performed meta-analysis–based projection modeling using pooled overall survival hazard ratios reported in a recent systematic review and meta-analysis of 4,022 NPM1-mutated AML patients. For the ELN-favorable subgroup, the pooled hazard ratio for overall survival associated with STMs (HR 1.95; 95% CI 1.39–2.73) was combined with externally derived baseline 5-year survival estimates under a proportional hazards framework. Absolute survival probabilities were calculated using established survival–hazard transformations, and results were translated into projected outcomes for a hypothetical cohort of 1,000 future ELN-favorable NPM1-mutated AML patients. Uncertainty was quantified using the meta-analytic confidence interval. Results: Assuming a baseline 5-year overall survival of 70% for ELN-favorable NPM1-mutated AML patients without STMs, the presence of STMs was projected to reduce 5-year survival to approximately 50%. Based on the confidence interval of the pooled hazard ratio, projected survival with STMs ranged from 38% to 61%. In a simulated cohort of 1,000 patients, this corresponds to approximately 300 deaths at 5 years in the absence of STMs versus 501 deaths in the presence of STMs, yielding an estimated 200 excess deaths per 1,000 patients (plausible range: 90–320 excess deaths). These findings were robust across visualizations including absolute survival projections, stacked outcome distributions, and excess mortality estimates. Conclusions: Meta-analysis–based projection modeling demonstrates that secondary-type mutations may negate much of the favorable prognosis traditionally attributed to ELN-favorable NPM1-mutated AML, translating into a large absolute survival decrement in future patient cohorts. Converting relative risk estimates into absolute survival and event projections provides clinically actionable insight and supports the incorporation of secondary-type mutation status into risk stratification, trial design, and therapeutic decision-making.

Background Documentation burden contributes significantly to psychiatric clinician burnout, with psychiatrists spending an average of three hours per workday on administrative tasks rather than direct patient care. While quantitative studies demonstrate that ambient artificial intelligence (AI) scribes reduce workload and improve documentation quality, the experiential perspectives of clinicians and standardized patients (SPs) remain unexplored, particularly in psychiatry where the therapeutic relationship is central to practice. Our aims were to explore clinician and standardized simulated patient perspectives on ambient AI scribes in psychiatric consultations, examining experiences, concerns, and conditions for successful implementation. Methods This study was part of a cross-sectional quantitative simulation study, where qualitative part was planned a-priori. We conducted focus groups with psychiatrists and SPs following a simulation-based crossover study comparing traditional versus AI-assisted documentation. Data were analyzed using Braun and Clarke’s reflexive thematic analysis framework. Results In this focus group session with eight psychiatrists and four SPs, total six themes were identified: (1) Clinical Presence, whereby reduced documentation barriers were perceived to allow more authentic human connection; (2) Liberation from documentation burden, with clinicians describing cognitive and emotional relief beneficial for their mental health; (3) Clinical Competence Amplified, in which clinicians perceived the AI scribe as enhancing practice through intelligent translation to psychiatric terminology and prompting for missed assessments; (4) Calibrating Trust, where participants navigated initial uncertainty while recognizing the importance of human oversight; (5) Privacy, stigma, and psychiatric exceptionalism, highlighting unique consent and confidentiality considerations for psychiatric populations; and (6) Perceptions of an enhanced standard of care, with both groups expressing strong preference for AI-assisted consultations with conditions for adoption. Conclusion In this exploratory simulation-based qualitative study, AI scribes were perceived as having potential to support more present, patient-centered psychiatric consultations by reducing documentation burden. Implementation should be cautious and requires transparent consent processes, clinician training in behavioral adaptations, robust human oversight, specialty-specific templates, and real-world evaluation with patients receiving psychiatric care.

ABSTRACT Objective To evaluate intracochlear electrocochleography (ECochG) amplitude parameters during cochlear implantation (CI) using a novel multi‐frequency ECochG algorithm. Methods A multi‐institutional, prospective cohort study was performed at 18 high‐volume CI centers. The inclusion criteria were adults with sensorineural hearing loss and audiometric thresholds of ≤ 90 dB hearing level at 500 Hz undergoing CI with Advanced Bionics (Valencia, CA) Ultra 3D devices between 2024 and 2025. ECochG recordings were performed with simultaneous multi‐frequency stimulation of four frequencies between 125 and 4000 Hz during cochlear implant insertion. Concurrent multi‐frequency recording allowed extraction of amplitude and phase of each frequency individually. Post hoc analysis was performed to determine the difference in the number of drop alarms between single‐ and multi‐frequency ECochG. An ECochG amplitude drop of 6 dB was defined as a drop alarm. Insertion track patterns were compared between single‐ and multi‐frequency ECochG. Results One hundred ninety‐five ears were included. Mean number of drop alarms for the single‐frequency algorithm was 1.72 (95% CI: 1.52, 1.92; median 1) compared to 0.42 (95% CI: 0.31, 0.53; median 0) for multi‐frequency; p < 0.001. The number of Type C patterns (rise in amplitude during insertion followed by a drop) decreased with the multi‐frequency ECochG algorithm compared to the single‐frequency ECochG algorithm. The number of Type D patterns (no‐response) decreased, indicating that multi‐frequency ECochG generated more responses across the cochlea than single‐frequency ECochG. Conclusions A novel multi‐frequency ECochG algorithm during CI is associated with fewer drop alarms and altered insertion track patterns, which may provide a more accurate assessment of the cochlear microenvironment. Level of Evidence 3.

Background: Agentic artificial intelligence (AI) systems, characterized by autonomous goal-directed behavior, multi-step reasoning, task decomposition, and tool use, are increasingly proposed for healthcare applications. However, their autonomy raises concerns regarding transparency, accountability, and human oversight. While explainable AI (XAI) has been widely studied in traditional predictive models, less is known about how explainability is implemented within agentic architectures. Objective: To map the emerging literature on explainable agentic AI (XAAI) in healthcare and characterize the types, scope, and forms of explainability used in these systems. Methods: A scoping review was conducted following PRISMA-ScR guidelines. PubMed, Embase, IEEE Xplore, and ACM Digital Library were searched through November 2025. Eligible studies described healthcare-related agentic AI systems incorporating explicit explainability mechanisms. Data were extracted on system architecture, explainability type (intrinsic, post hoc, hybrid), explanation scope (local, global), explanation form, and reported clinical outcomes. Results: Nine studies met the inclusion criteria. All systems demonstrated core agentic features, including autonomy, task decomposition, and tool integration, often within multi-agent frameworks. Explainability was predominantly intrinsic and workflow-native, typically delivered through textual reasoning traces and example-based grounding in retrieved clinical evidence. Feature-based and global explanations were comparatively rare and largely confined to hybrid architectures. Across domains including radiology, neurology, psychiatry, and biomedical research, XAAI systems were reported to improve performance and interpretability relative to baseline models in the included studies. However, these findings were derived from heterogeneous, predominantly experimental or retrospective studies, and structured human-in-the-loop oversight was infrequently described. Conclusions: Current XAAI systems appear to emphasize process transparency and evidence grounding rather than mechanistic model-level attribution. The available evidence remains limited and heterogeneous, and findings should be interpreted as early trends rather than established characteristics. Further progress will require standardized evaluation frameworks, clearer reporting of oversight mechanisms, and validation in real-world clinical settings to support safe and trustworthy integration of agentic AI into healthcare practice.

Background: Large language models (LLMs) are being used by surgeons for education and reference yet concerns about hallucinations and reliability limit safe adoption. Retrieval-augmented generation (RAG) can offer a potential solution by grounding responses in a high-quality external database (e.g., medical textbooks) to enhance accuracy. However, performance tradeoffs across different RAG configurations—many of which exponentially increase computational cost—remain poorly characterized. Methods: In total, 120 lightweight, open-source RAG configurations were evaluated across 40 plastic surgery-focused question-answering tasks (20 single-hop, 20 multi-hop), spanning multiple subspecialties (4800 total evaluations). Configurations varied by base LLM (Phi-3-mini-128k-instruct vs. BioMistral-7B), embedding model, database size, chunk size, and query hop type. Performance was assessed using semantic similarity (Ragas) to physician-validated reference answers. Performance was analyzed using linear mixed-effects regression with query as a random effect and fixed and interaction effects selected via likelihood testing and AIC. Results: High performance was achievable using lightweight, open-source models. While BioMistral-7B had high mean sematic similarity under specific configurations (mean semantic similarity up to 0.786), Phi-3-mini-128k-instruct demonstrated more consistent performance across query complexity. Larger database sizes significantly improved semantic similarity, with the largest gain at intermediate sizes (e.g., size 5: +0.043, p = 0.001). Embedding choice had a strong effect, with bge-large-en-v1.5 improving performance (p = 0.0016) and Bio_ClinicalBERT markedly reducing it (p < 0.001). Multi-hop queries substantially reduced performance (p < 0.001), though this effect was attenuated for Phi-3-mini-128k-instruct via a strong model × hop-type interaction (p < 0.001). Conclusions: RAG systems for plastic surgery do not require large proprietary models, as performance depends on configuration choices and interaction effects rather than isolated components. With advancements, predictive modeling may enable resource-efficient, safe deployment of clinical RAG systems.