Citations

PURPOSE Large language models (LLMs) have demonstrated remarkable versatility in oncology applications, such as cancer staging and survival analysis. Despite their potential, ethical concerns such as data privacy breaches, bias in training data, lack of transparency, and risks associated with erroneous outputs pose significant challenges to their adoption in high-stakes oncology settings. Therefore, we aim to explore the ethical challenges associated with LLM-based applications in oncology and evaluate emerging techniques designed to address these issues. METHODS Following the Preferred Reporting Items for Systematic Reviews and Meta-Analyses framework, a systematic review was conducted to evaluate publications related to ethical issues of LLMs in oncology across eight academic databases (eg, PubMed, Web of Science, and Embase) between January 1, 2019, and December 31, 2024. RESULTS The search retrieved 4,319 published articles, of which 65 publications were preserved and included in our analysis. We identified six prevalent ethical challenges in oncology, including trust, equity, privacy, transparency, nonmaleficence, and accountability. We then evaluated emerging technical solutions to mitigate ethical challenges and summarized evaluation metrics used to assess these solutions' effectiveness. CONCLUSION This review provides actionable recommendations for responsibly deploying LLMs in oncology, ensuring adherence to ethical guidelines, and fostering improved patient outcomes. By bridging technical and clinical perspectives, this review offers a foundational framework for advancing ethical artificial intelligence applications in oncology and highlights areas for future research.

Abstract Background Lung adenocarcinoma in never-smokers is a major public health burden, especially among East Asian women. Polygenic risk scores (PRSs) are promising for risk stratification but are primarily developed in European-ancestry populations. We aimed to develop and validate single- and multi-ancestry PRSs for East Asian never-smokers to improve lung adenocarcinoma risk prediction. Methods PRSs were developed using genome-wide association study summary statistics from East Asian (8002 cases; 20 782 controls) and European (2058 cases; 5575 controls) populations. Single-ancestry models included a PRS with 25 single nucleotide variants (PRS-25), PRS clumping and thresholding (CT), and LDpred2; multi-ancestry models included LDpred2 plus European PRS-128, PRS-CSx, and clumping and thresholding, super learning, empirical Bayes (CT-SLEB). Performance was evaluated in independent East Asian data from the Female Lung Cancer Consortium and externally validated in the Nanjing Lung Cancer Cohort. We assessed predictive accuracy via area under the curve operating procedure (AUC), with 10-year and (ages 30-80 years) absolute risks estimates. Results The best multi-ancestry PRS, using East Asian and European data via CT-SLEB, outperformed the best East Asian-only (LDpred2; AUC = 0.629, 95% confidence interval [CI] = 0.618 to 0.641), achieving an AUC of 0.640 (95% CI = 0.629 to 0.653) and odds ratio of 1.71 (95% CI = 1.61 to 1.82) per standard deviation increase. Nanjing Lung Cancer Cohort validation confirmed robust performance (AUC = 0.649, 95% CI = 0.623 to 0.676). The top 20% PRS group had a 3.92-fold higher lung adenocarcinoma risk than the bottom 20%. Further, the top 5% PRS group reached a 6.69% lifetime absolute risk. Notably, this group reached the average population 10-year lung adenocarcinoma risk at age 50 years (0.42%) by age 41 years, 9 years earlier. Conclusions Multi-ancestry PRS approaches enhance lung adenocarcinoma risk stratification in East Asian never-smokers, with consistent external validation, suggesting future clinical utility.

Abstract Error analysis is a critical step in evaluating and improving clinical concept extraction models, the most common clinical natural language processing (NLP) task. Unlike corpus annotation, which follows standardized protocols for creating gold-standard datasets, error analysis requires nuanced judgment grounded in both clinical expertise and NLP knowledge. This is especially important given the heterogeneity of clinical text, where variations in documentation style, note structure, and terminology can substantially influence model behavior. Despite its importance, there is currently no standardized, user-level framework to support systematic error analysis in clinical concept extraction task. In this study, we developed and validated MedError, a machine-assisted, human-in-the-loop framework designed to standardize and enhance error analysis for clinical concept extraction tasks. We collected and manually curated a corpus of 1,187 unique errors from a total of 4,237 notes across three different distinct hospitals. The error categories were defined using our previously validated error taxonomy and included 480 false negatives and 707 false positives across 25 error types and 48 clinical concept categories. We evaluated the performance of three proprietary and three open-source large language models (LLMs) in automatically classifying these errors into 26 and 15 predefined categories. We further developed a machine-assisted framework, MedError, which integrates best practices in error analysis, LLM-assisted classification and reasoning, and a user-friendly interface to enable more efficient, reproducible, and context-aware error analysis. The framework supports both single-site and federated multisite error analysis, facilitating the effective translation of clinical NLP systems into real-world settings.

ABSTRACT Background & Aims Steatotic liver disease (SLD) is characterised by liver fat accumulation exceeding 5%. Lean body weight (BMI ≤ 25 kg/m 2 ) with non‐alcoholic SLD is a rare phenotype, and the balance of risks related to metabolic conditions or genetic predisposition, or the risk of progressive liver disease, is not known. This study evaluates clinical and genetic predictors of non‐alcoholic SLD and advanced liver disease in lean individuals. Methods We used International Classification of Disease codes, radiology and pathology review to identify 177 lean non‐alcoholic SLD cases (47 cryptogenic, 130 MASLD), 677 matched lean controls, and 3090 overweight/obese SLD cases in the Mayo Clinic Biobank and Tapestry databases. We performed case–control and cross‐sectional comparisons of clinical and genetic factors between these groups, using univariable and multivariable analyses. Results Lean individuals with non‐alcoholic SLD exhibited metabolic and genetic profiles that were intermediate between those of lean controls without SLD and overweight/obese SLD individuals, including intermediate rates of diabetes, hypertension, hyperlipidaemia and of homozygosity for the risk allele of PNPLA3 . Multivariable analysis indicated that diabetes was an independent predictor of SLD among lean individuals. Among lean individuals with non‐alcoholic SLD, those with metabolic‐associated SLD (MASLD), as compared to those with cryptogenic SLD (without metabolic risk factor), were more likely to be homozygous for risk alleles in GCKR . The Fib‐4 score, a tool for screening advanced liver disease in SLD, accurately predicted advanced fibrosis among lean individuals with non‐alcoholic SLD. Conclusion Diabetes serves as a primary predictor of non‐alcoholic SLD in lean individuals. These results support the recommendation to screen for SLD in patients with diabetes, regardless of BMI. The GCKR risk allele is associated with MASLD in lean individuals, and the risk factors for cryptogenic SLD remain unclear.

Abstract Frailty is highly prevalent in cardiovascular disease and associates with poor clinical outcomes. Several cardiovascular conditions benefit from invasive interventions, but the impact of these procedures in frail patients remains unclear. We thus performed a systematic review and meta-analysis to summarize evidence and quantify the pooled effect of clinical outcomes of invasive cardiovascular interventions among frail vs. non-frail patients. Systematic search of two electronic databases from January 2015 to April 2025 was performed including all randomized controlled trials (RCTs) and prospective cohort studies examining outcomes of invasive cardiovascular interventions stratified by frailty. Outcomes included all-cause mortality, hospitalizations, and other adverse outcomes. Risk ratios (RRs) and their 95% confidence intervals (CI) were assessed using random-effects models. Thirty-one studies (10 RCTs, 21 prospective, 3 retrospective studies) with a total of 27 534 participants were included. Compared with non-frail patients, frail patients after cardiovascular procedures were at higher risks of all-cause mortality (RR 3.14, 95% CI: 2.29-4.30), and hospitalizations (RR 1.88, 95% CI: 1.29-2.73). Frail patients undergoing transcatheter aortic-valve implantation and percutaneous mitral-valve interventions did not have significant increased mortality risks compared with non-frail patients (P > 0.05). Pre-frailty predicted lower mortality risks amongst patients undergoing transcatheter vs. surgical aortic valve replacement (HR 0.78, 95% CI: 0.67-0.92). Frail patients with non-ST elevation myocardial infarction derived similar mortality benefits from an invasive vs. conservative approach (HR 1.13, 95% CI: 0.64-2.01). In conclusion, frailty prognosticates adverse outcomes following invasive cardiovascular interventions and influences treatment efficacy, highlighting its significance in treatment selection

Abstract Background Dose‐driven continuous scanning (DDCS) enhances the efficiency and precision of proton pencil beam delivery by reducing beam pauses inherent in discrete spot scanning (DSS). However, current DDCS optimization studies using traveling salesman problem (TSP) formulations often rely on fixed beam intensity and computationally expensive interpolation for move spot generation, limiting efficiency and methodological robustness. Purpose This study introduces a Break Spot‐Guided (BSG) method, combined with two acceleration strategies—dose rate skipping and bounding—to optimize beam intensity while minimizing beam delivery time (BDT). In addition, a ‐method is proposed for efficient move spot generation, aiming to balance computational efficiency and dose calculation accuracy. Methods The BSG framework simplifies the original mixed‐integer nonlinear programming (MINLP) problem by transforming it into a series of TSP evaluations using break‐spot dose rates. The two acceleration strategies are integrated to reduce computational complexity. The ‐method is introduced to efficiently generate move spots without compromising dose fidelity. The performance of the BSG method, with and without the acceleration strategies, and the ‐method, was evaluated across multiple clinical cases. Results By jointly optimizing the scan path and dose rate, the proposed BSG‐S‐B method offered the most efficient delivery, achieving a 64% reduction in BDT compared to DSS and a 57% reduction compared to Liu's method in the prostate case. These gains were consistently observed across all cases. Utilizing acceleration strategies, BSG‐S‐B significantly reduced computation time from hours to minutes (e.g., from 10 525.7 to 108.5 s in the prostate case, and from 6202.4 to 129.8 s in the head‐and‐neck case). When combined with the ‐method, it reduced the number of move spots by more than 50%, while still ensuring dosimetric quality. Conclusions The proposed framework significantly improves computational efficiency, making scan path optimization with dose rate selection feasible for clinical DDCS delivery. The ‐method effectively reduces computational complexity while maintaining dose calculation accuracy, demonstrating strong potential for clinical adoption in proton therapy treatment planning.

Abstract Repetitive head impacts are associated with structural brain changes and an increased risk for chronic traumatic encephalopathy, a progressive neurodegenerative disease that can only be diagnosed after death. Chronic traumatic encephalopathy is defined by the abnormal accumulation of phosphorylated tau protein, particularly at the depths of the superior frontal sulci, suggesting that sulcal morphology may serve as a relevant structural biomarker. Contact sport athletes, such as former football players, are at elevated risk due to their prolonged exposure to repetitive head impacts. Cortical atrophy linked to underlying tau accumulation may result in shallower and wider sulci, potentially making sulcal morphology an imaging marker for identifying individuals at risk for this disease. This study investigated sulcal morphological differences in former football players and examined associations with age, football-related exposure, clinical diagnosis of traumatic encephalopathy syndrome, levels of certainty for chronic traumatic encephalopathy pathology, neuropsychological performance, and positron emission tomography imaging using flortaucipir. We analysed structural magnetic resonance imaging data from 169 male former football players (mean age 57.2 (8.2) years, range 45–74) and 54 age-matched, unexposed asymptomatic male controls (mean age 59.4 (8.5) years, range 45–74). Sulcal depth and width were quantified using the CalcSulc, focusing on two regions in each hemisphere commonly affected by chronic traumatic encephalopathy pathology: the superior frontal and occipitotemporal sulci. Generalized least squares models were used to assess group differences and interactions with age and football exposure variables, including age of first exposure, total years played, and cumulative head impact exposure. An analysis of covariance evaluated relationships between sulcal morphology, clinical measures, and flortaucipir uptake, adjusting for age, race, body mass index, education, imaging site, apolipoprotein E4 status, and total intracranial volume. Former football players demonstrated significantly shallower sulcal depth in the left superior frontal sulcus compared to unexposed controls. Earlier age of first exposure and longer football careers were associated with greater widening of the left occipitotemporal sulcus. Higher cumulative head impact exposure was linked to reduced sulcal depth in the left superior frontal region. However, sulcal morphology was not associated with clinical diagnosis, levels of certainty, neuropsychological test performance, or flortaucipir imaging. These findings suggest that sulcal morphology may reflect cumulative exposure to repetitive head impacts, particularly in brain regions vulnerable to chronic traumatic encephalopathy pathology. Future ante- and post-mortem validation studies are needed to determine whether sulcal morphology can serve as a reliable in vivo biomarker of risk.

Abstract Background Physical activity (PA) is an established protective factor for colorectal cancer (CRC), but it is unclear if genetic variants modify this effect. To investigate this possibility, we conducted a genome-wide gene–PA interaction analysis. Methods Using logistic regression and two-step and joint tests, we analyzed interactions between common genetic variants across the genome and PA in relation to CRC risk. Self-reported PA levels were categorized as active (≥ 8.75 MET-h/wk) vs. inactive (< 8.75 MET-h/wk) and as study- and sex-specific quartiles of activity. Results PA had an overall protective effect on CRC (OR [active vs. inactive] = 0.85; 95%CI = 0.81–0.90). The two-step GxE method identified an interaction between rs4779584, an intergenic variant near the GREM1 and SCG5 genes, and PA for CRC risk (p-interaction = 2.6×10 − 8 ). Stratification by genotype at this locus showed a significant reduction in CRC risk by 20% in active vs. inactive participants with the CC genotype (OR = 0.80; 95%CI = 0.75–0.85), but no significant PA–CRC association among CT or TT carriers. When PA was modeled as quartiles, the 1-d.f. GxE test identified that rs56906466, an intergenic variant near the KCNG1 gene, modified the association between PA and CRC (p-interaction = 3.5×10 − 8 ). Stratification at this locus showed that increase in PA (highest vs. lowest quartile) was associated with a lower CRC risk solely among TT carriers (OR = 0.77; 95%CI = 0.72–0.82). Conclusions In summary, we identified two genetic variants that modified the association between PA and CRC risk. One of them, related to GREM1 and SCG5 , suggests that the bone morphogenetic protein (BMP)-related, inflammatory, and/or insulin signaling pathways may be associated with the protective influence of PA on colorectal carcinogenesis.