Sangzin Ahn

Background Decision models return a probability for each predefined answer option instead of generating text. If these probabilities identify unreliable answers, those answers could be referred to a clinician or a stronger model while the rest are accepted. Objective To evaluate whether the probabilities of three decision models identify their errors on medical licensing questions and what referral of the least confident answers achieves. Methods This exploratory in silico study used all 435 KorMedMCQA doctor test questions and all 1,273 MedQA test questions. Jev, OpenAI Decisions and Clef were compared with answer-letter probabilities from GPT-6 Luna, the generative model on which OpenAI Decisions is built. Each model's least confident 20% of answers were referred to GPT-6.1 Sol, and the result was compared with random referral and with an oracle that knows which answers are wrong. Rotation of the answer options tested the stability of referral decisions. Results The decision models answered 83.4% to 92.0% of Korean and 84.8% to 87.7% of English questions correctly, with error-detection areas under the receiver operating characteristic curve (AUROCs) of 0.856 to 0.935 and 0.864 to 0.903. Luna letter probabilities were exactly one for most answers, and OpenAI Decisions ranked errors better than the letter probabilities of its base model, by 0.091 (95% CI 0.007 to 0.180) in Korean and 0.104 (0.062 to 0.147) in English. The least confident 20% of answers held 56% to 91% of Korean and 62% to 72% of English errors. Referring them to GPT-6.1 Sol raised accuracy by 6 to 9 percentage points, achieving 49% to 91% of the gain available to an oracle, at a quarter to a third of Sol's cost. Rotating options changed 7% to 16% of referral decisions. In English, errors made with high confidence were stable under rotation and often shared by other models. Conclusions Decision-model probabilities identified most errors and allowed a fixed share of answers to be referred to a stronger model at a fraction of its cost, whereas a generative model's letter probabilities were mostly saturated. Confident errors and order-dependent referral remained, and prospective evaluation with clinicians is needed.

Artificial intelligence (AI) agents extend large language models from single-turn text generation to systems that pursue goals through planning, retrieval, tool use, code execution, memory, feedback, and role coordination. In medicine and biomedical research, this shift is creating early systems for clinical calculations, risk prediction, oncology decision support, omics analysis, hypothesis development, laboratory automation, and research writing. However, the evidence remains uneven. Clinical examples are the most defensible when agents use validated calculators, curated clinical tools, or guideline-grounded modules under human oversight. Biomedical discovery systems exhibit broader workflow capabilities; however, many claims still rely on preprints, narrow benchmarks, simulated settings, or domain-specific demonstrations. For clinicians and biomedical researchers, the immediate challenge is not to decide whether agents will replace experts but to understand what tasks can be delegated, what evidence is needed, and what human judgment must be preserved. This narrative review explains what makes an AI system agentic, summarizes its representative clinical and discovery applications, and outlines safeguards for evaluation, reproducibility, and oversight. Biomedical readers should expect AI agents to enter medicine and research first as constrained, auditable workflow infrastructures. These infrastructures may reorganize biomedical work; however, accountability should remain with the clinicians and investigators.

Background: Disability weights (DWs) quantify the severity of health loss and are essential for estimating disability-adjusted life years in the Global Burden of Disease (GBD) framework. Conventional DW estimation relies on resource-intensive population surveys that are difficult to update or adapt to emerging health states. Large language models (LLMs) may offer a scalable alternative by approximating human perceptions of disease severity through structured judgment tasks. Methods: This exploratory study evaluated the alignment between LLM-derived and human-derived DW rankings using 222 health states from GBD 2010. All possible pairwise comparisons (24,531 pairs, each repeated three times) were conducted across four LLMs (GPT-5 mini, GPT-5, Claude Haiku 4.5, and Claude Sonnet 4.5). DWs were estimated via probit regression and evaluated using Spearman’s rank correlation and Steiger’s z test. The effects of prompt language (English vs. Korean), cultural role prompting, and medical specialist role prompting on alignment were examined. Additionally, the Binomial-Logit Indifference-Point (BLIP) estimator was proposed and validated through leave-one-out cross-validation for estimating DWs for health states without established values. Results: All four LLMs showed high rank correlation with GBD 2010 DWs (Spearman’s ρ = 0.893 to 0.909), with no significant inter-model differences. Korean-language prompting significantly improved alignment with Korean DWs (ρ = 0.756 vs. 0.715, p = 0.011), and Korean cultural role prompting improved alignment with both GBD 2010 DWs (ρ = 0.922 vs. 0.909, p = 0.002) and Korean DWs (ρ = 0.738 vs. 0.715, p = 0.001). Medical specialist role prompting significantly reduced alignment with GBD 2010 DWs (ρ = 0.895 vs. 0.909, p = 0.001). BLIP demonstrated strong agreement with GBD 2010 DWs (Pearson’s r = 0.862, MAE = 0.066) and produced plausible estimates for Long COVID (mild: 0.020, moderate: 0.298, severe: 0.529). Conclusions: LLMs can approximate human perceptions of disease severity with high rank-order consistency. Prompt language and role framing significantly influenced alignment, with culturally grounded lay prompting enhancing and specialist prompting reducing correspondence with population-based DWs. BLIP provides a practical framework for generating provisional DW estimates for emerging or underrepresented health states when conventional surveys are infeasible.

Background/Objectives: Macrophages polarized into M1 and M2 phenotypes differentially regulate immune and drug responses. Despite their distinct functional roles, differences in UDP-glucuronosyltransferase (UGT) expression and enzymatic activity between M1 and M2 macrophages remain poorly understood. This study aimed to characterize differential UGT expression in M1 and M2 macrophages and to elucidate how UGT-mediated prostaglandin E₂ (PGE₂) glucuronidation modulates macrophage inflammatory responses. Methods: THP-1 cells were chemically differentiated into macrophages (M0) and subsequently polarized into M1 and M2 phenotypes. UGT expression profiles were assessed using RT-PCR, quantitative RT-PCR (qRT-PCR), and Western blot. UGT activity was compared by quantifying glucuronide metabolites derived from UGT-specific substrates using LC-MS/MS, along with measurement of free PGE₂ and PGE₂-glucuronide by ELISA. Pro-inflammatory cytokine expression and secretion in M1 macrophages were quantified using qRT-PCR and ELISA. Results: Expression of UGT1A1, UGT1A4, UGT1A5, UGT1A9, and UGT2B7 were markedly higher in M1 compared with M2 macrophages at both the mRNA and protein levels. Enhanced UGT activity in M1 macrophages was reflected by increased formation of estradiol-3-glucuronide and naloxone-3-glucuronide (both p < 0.01) and was attenuated in a concentration-dependent manner by diclofenac. Furthermore, PGE₂ glucuronidation was more pronounced in M1 macrophages, and inhibition of UGTs with atazanavir reduced PGE₂-glucuronide formation and pro-inflammatory cytokine production, including IL-1β, IL-6, and TNF-α. Conclusion: UGT-mediated PGE₂ glucuronidation in M1 macrophages contributes to the regulation of pro-inflammatory cytokine production. Collectively, these findings support a role for UGTs as modulators of inflammatory signaling, with differential expression and activity between M1 and M2 macrophages.

Purpose: To develop and evaluate a large language model (LLM)-based learning tool, featuring virtual patients (VPs) and virtual assessors (VAs), and to assess its impact on medical students’ perceptions of history-taking education compared to conventional learning methods.Methods: A tool using the GPT-4 API was developed to provide seven clinical VP scenarios and a VA that delivered both immediate, reflective dialogue and comprehensive written feedback. First- and second-year medical students participated in a 6-day study. Pre- and post-participation surveys using a 5-point Likert scale assessed perceptions of the LLM tool versus conventional methods across usability, self-efficacy, and feedback quality domains.Results: Twenty-one students completed the study. The LLM-based tool demonstrated statistically significant improvements over conventional methods in all assessed domains. Students reported greater comfort during practice (mean 4.57 vs. 2.95, p=0.0002). Significant gains were seen in six of eight self-efficacy measures, including confidence in handling unfamiliar cases (4.00 vs. 2.90, p=0.0002). All nine feedback quality dimensions improved significantly, with feedback perceived as more specific (4.43 vs. 3.24, p=0.0005) and personalized (4.19 vs. 3.19, p=0.0001).Conclusion: An LLM-based learning tool featuring VPs and VAs can significantly enhance medical students’ perceived learning experience in history-taking education. It offers a scalable, accessible, and cost-effective complementary training method. Future research should validate these subjective improvements with objective performance metrics.

Background/Objectives: Macrophages polarized into M1 and M2 phenotypes differentially regulate immune and drug responses. Despite their distinct functional roles, differences in UDP-glucuronosyltransferase (UGT) expression and activity between M1 and M2 macrophages remain poorly understood. This study aimed to characterize differential UGT expression in M1 and M2 macrophages and to elucidate how UGT-mediated prostaglandin E₂ (PGE₂) glucuronidation modulates macrophage inflammatory responses. Methods: THP-1 cells were chemically differentiated into macrophages (M0) and subsequently polarized into M1 and M2 phenotypes. UGT expression profiles were assessed using RT-PCR, quantitative RT-PCR, and Western blotting. UGT enzymatic activity was compared by quantifying glucuronide metabolites derived from UGT-specific substrates using LC-MS/MS, along with measurement of free PGE₂ and PGE₂-glucuronide by ELISA. Pro-inflammatory cytokine expression and secretion in M1 macrophages were quantified using quantitative RT-PCR and ELISA. Results: UGT1A1, UGT1A4, UGT1A5, UGT1A9, and UGT2B7 were markedly upregulated in M1 compared with M2 macrophages at both the mRNA and protein levels. UGT enzymatic activity was significantly higher in M1 macrophages (p < 0.01) and was attenuated in a concentration-dependent manner by the UGT inhibitor diclofenac. Enhanced UGT activity in M1 macrophages was reflected by increased formation of estradiol-3-glucuronide and naloxone-3-glucuronide (both p < 0.01). Furthermore, PGE₂ glucuronidation was more pronounced in M1 macrophages, and inhibition of UGTs with atazanavir reduced PGE₂-glucuronide formation and pro-inflammatory cytokine production, including IL-1β, IL-6, and TNF-α. Conclusion: UGT-mediated PGE₂ glucuronidation in M1 macrophages plays a critical role in sustaining pro-inflammatory cytokine responses. These findings identify UGTs as important regulators of macrophage phenotype and inflammatory signaling.

Rapid advancements in large language models (LLMs) have fundamentally transformed research practices across academic disciplines, with considerable adoption rates among researchers. While empirical studies have demonstrated the substantial positive effects of LLMs on learning outcomes and cognitive performance, these technological advances present a paradoxical challenge in maintaining critical thinking capabilities. LLMs offer unprecedented efficiency in research tasks in literature reviews, analysis, and writing, by significantly reducing task completion time while improving output quality. However, this efficiency stems largely from cognitive offloading, which is the delegation of mental processes to external systems, raising concerns about the potential weakening of human analytical abilities. Cognitive load theory provides a framework for distinguishing between the beneficial reduction of unnecessary cognitive burden and problematic offloading of essential cognitive processes required for deep understanding and critical analysis. Experimental evidence suggests that while LLM users experience reduced cognitive load across multiple dimensions, their critical reasoning performance may suffer compared with traditional search methods. The fundamental challenge lies in balancing the efficiency gains of LLM integration while preserving rigorous analytical thinking. Medical researchers must develop strategic approaches that leverage LLM capabilities while maintaining active engagement with primary sources and complex reasoning tasks. Success lies in recognizing that traditional research methods may represent essential investments in preserving critical thinking skills, suggesting LLM integration involves selective application rather than wholesale cognitive offloading.