BACKGROUND:Severe fever with thrombocytopenia syndrome (SFTS) is a life-threatening tick-borne disease endemic to East Asia. Despite its high fatality rate and increasing incidence, the long-term kinetics of humoral immune responses in patients with SFTS remain poorly understood. METHODS:We analyzed serum from 109 patients hospitalized with SFTS between 2015 and 2023. Antibody responses were assessed using indirect immunofluorescence assay (IFA), enzyme-linked immunosorbent assay (ELISA), and plaque reduction neutralization test (PRNT). Antibody kinetics were described by outcome, but time-to-seroconversion was not compared because follow-up differed before death. RESULTS:The SFTSV viral load peaked within 1-9 days after symptom onset, before declining. SFTSV-specific IgG antibodies were detected by IFA from days 5-9 and by ELISA from days 10-14, peaking on days 20-24. IgM responses appeared later, peaking on days 20-24 (IFA) and 25-29 (ELISA). Neutralizing antibodies (PRNT50 ≥ 1:10) emerged from day 10 and peaked on days 20-24. SFTSV-specific IgG and neutralizing antibodies remained detectable for up to 5 years, with all patients evaluated at the 5-year follow-up remaining seropositive (3/3). Seroconversion of IgG and neutralizing antibodies occurred only in survivors and was absent in fatal cases. IgM responses were completely undetectable in fatal patients. CONCLUSION:SFTSV-specific binding and neutralizing antibodies developed strongly in survivors and persisted for up to 5 years. Seroconversion was not detected in fatal cases, although this may reflect limited follow-up. These findings indicate a link between early humoral responses and clinical outcomes, but do not establish protective immunity.

Wearable devices generate continuous physiological and behavioral data, but converting these signals into clinically reviewable biomarker hypotheses remains labor-intensive. We introduce CoDaS, an AI co-data-scientist that integrates multi-agent hypothesis generation, deterministic statistical analysis, adversarial validation and literature-grounded interpretation under human oversight. Across three wearable cohorts comprising 9,279 participant-observations, CoDaS prioritized candidate associations for mental-health and metabolic endpoints after internal checks for replication, stability, robustness and leakage. The system identified related circadian-instability signals associated with depression, including sleep-duration variability in DWB (ρ = 0.252, p < 0.001) and sleep-onset variability in GLOBEM (ρ = 0.126, p < 0.001), and derived a wearable cardiovascular-fitness index associated with insulin resistance (steps/resting heart rate; ρ = -0.374, p < 0.001). Adding these features to demographic models produced modest gains (ΔR^2 = 0.040 for depression, 0.021 for insulin resistance). In a 12-clinician review totaling approximately 25 active hours, clinician validity judgments aligned with CoDaS confidence tiers (ρ = 0.67, p = 0.005), whereas added clinical value and confidence to act were rated lower. CoDaS supports traceable, hypothesis-generating prioritization of wearable candidate biomarkers.

We aimed to characterize human immunodeficiency virus (HIV) drug-resistance mutations among patients in Korea HIV/acquired immunodeficiency syndrome (AIDS) Cohort. Among 1,579 surveyed between December 2006 and December 2020, 148 had resistance mutations (64 ART-naïve and 84 ART-experienced). The most common mutations in ART-naïve patients were related to non-nucleoside reverse transcriptase inhibitors (V179F/D/E/L; 50.0%), whereas those in ART-experienced were M184V/I (73.8%), conferring resistance to nucleotide reverse transcriptase inhibitors. Multivariate logistic regression identified poor compliance (adjusted odds ratio [aOR], 18.780; 95% confidence interval [CI], 1.883-187.300; P=0.012) and viral load ≥1-100,000 copies/mL at ART initiation (aOR, 3.117; 95% CI, 1.160-8.374; P=0.024) as being associated with resistance mutations.

Hallucinations in foundation models arise from autoregressive training objectives that prioritize token-likelihood optimization over epistemic accuracy, fostering overconfidence and poorly calibrated uncertainty. We define medical hallucination as any model-generated output that is factually incorrect, logically inconsistent, or unsupported by authoritative clinical evidence in ways that could alter clinical decisions. We evaluated 11 foundation models (7 general-purpose, 4 medical-specialized) across seven medical hallucination tasks spanning medical reasoning and biomedical information retrieval. General-purpose models achieved significantly higher proportions of hallucination-free responses than medical-specialized models (median: 76.6

Large Language Models (LLMs) as agents require careful behavioral adaptation. While adept at reactive tasks (e.g., medical reasoning), LLMs often struggle with proactive engagement, like unprompted identification of critical missing information or risks. We introduce **BehaviorBench**, a comprehensive dataset to evaluate agent behaviors across a clinical assistance spectrum. To rigorously test the current models, we also introduce **BehaviorBench-Hard**, a challenging subset where the performance of state-of-the-art models drops significantly, revealing weaknesses. To address these challenges, we propose **BehaviorSFT**, a novel training strategy using behavioral tokens to explicitly condition LLMs for dynamic behavioral selection which boosts performance on both benchmarks. Crucially, a blind clinician evaluation confirmed that our trained agents exhibit more realistic clinical behavior, striking a superior balance between helpful proactivity and necessary restraint versus standard fine-tuning or explicitly instructed agents. Project Page: https://behavior-adaptation.github.io/

Large repositories of image-caption pairs are essential for the development of vision-language models. However, these datasets are often extracted from noisy data scraped from the web, and contain many mislabeled instances. In order to improve the reliability of downstream models, it is important to identify and filter images with incorrect captions. However, beyond filtering based on image-caption embedding similarity, no prior works have proposed other methods to filter noisy multimodal data, or concretely assessed the impact of noisy captioning data on downstream training. In this work, we propose, theoretically justify, and empirically validate LEMoN, a method to identify label errors in image-caption datasets. Our method leverages the multimodal neighborhood of image-caption pairs in the latent space of contrastively pretrained multimodal models to automatically identify label errors. Through empirical evaluations across eight datasets and twelve baselines, we find that LEMoN outperforms the baselines by over 3% in label error detection, and that training on datasets filtered using our method improves downstream captioning performance by more than 2 BLEU points over noisy training.

Large language models (LLMs) deployed as agents introduce significant safety risks in clinical settings due to their potential for error and single points of failure. We introduce Tiered Agentic Oversight (TAO), a hierarchical multi-agent system that enhances AI safety through layered, automated supervision. Inspired by clinical hierarchies (e.g., nurse-physician-specialist) in hospital, TAO routes tasks to specialized agents based on complexity, creating a robust safety framework through automated inter- and intra-tier communication and role-playing. Crucially, this hierarchical structure functions as an effective error-correction mechanism, absorbing up to 24

Large Language Models (LLMs) have demonstrated remarkable performance on various medical question-answering (QA) benchmarks, including standardized medical exams. However, correct answers alone do not ensure correct logic, and models may reach accurate conclusions through flawed processes. In this study, we introduce the MedPAIR (Medical Dataset Comparing Physicians and AI Relevance Estimation and Question Answering) dataset to evaluate how physician trainees and LLMs prioritize relevant information when answering QA questions. We obtain annotations on 1,300 QA pairs from 36 physician trainees, labeling each sentence within the question components for relevance. We compare these relevance estimates to those for LLMs, and further evaluate the impact of these "relevant" subsets on downstream task performance for both physician trainees and LLMs. We find that LLMs are frequently not aligned with the content relevance estimates of physician trainees. After filtering out physician trainee-labeled irrelevant sentences, accuracy improves for both the trainees and the LLMs. All LLM and physician trainee-labeled data are available at: http://medpair.csail.mit.edu/.

BACKGROUND:The severe acute respiratory syndrome coronavirus 2 (SARS-CoV-2) continues to evolve, with each new strain becoming immunologically distinct from previous strains. The updated vaccine formulation for the 2023/24 season contained the mRNA of the XBB.1.5 sublineage. As the coronavirus disease-2019 (COVID-19) has become an endemic disease, a yearly assessment of updated COVID-19 vaccines is necessary. METHODS:This test-negative case-control study included patients who underwent COVID-19 testing between November 2023 and April 2024, and between July and August 2024, at eight university hospitals. To assess the change in VE over time, the study was divided into periods 1 (November-December 2023), 2 (January-April 2024), and 3 (July-August 2024) to assess the change of vaccine effectiveness (VE). VE (against laboratory-confirmed COVID-19 and COVID-19-related hospitalization) was calculated using adjusted odds ratios (OR) derived from multivariable logistic regression models. RESULTS:A total of 6198 patients were included (1671, 1811, and 2716 in period 1, 2, and 3, respectively), out of which 1115 (18.0%) underwent XBB.1.5 monovalent vaccination. The adjusted VE against laboratory-confirmed COVID-19 during each study period was 57.1% (95% confidence interval [CI]: 38.0-70.6, p < 0.001), 18.8% (95% CI -4.7-37.2, p = 0.110), and 3.3% (95% CI -15.4-19.1, p = 0.707), respectively. The corresponding statistics against COVID-19-related hospitalization were 74.0% (95% CI 40.7-78.6, p < 0.001), 34.2% (-6.8-46.4, p = 0.114), and 26.1% (4.7-42.8, p = 0.020), respectively. CONCLUSION:Despite waning VE against laboratory-confirmed COVID-19 over time, XBB.1.5 mRNA vaccines conferred sustained protection against COVID-19-related hospitalizations throughout the year.

Herein, we found that severe acute respiratory syndrome coronavirus 2 (SARS-CoV-2)-unexposed individuals exhibited an increased frequency of CD4+ T cells against SARS-CoV-2 membrane (M) protein, suggesting that SARS-CoV-2 M-reactive cells may be primed by previous infection with common cold coronaviruses (CCCoVs). We confirmed that CCCoV M-reactive CD4+ T cells cross-recognize SARS-CoV-2 M in unexposed individuals. Among coronavirus disease 2019 (COVID-19) convalescents and unexposed individuals, SARS-CoV-2 M-reactive CD4+ T cells exhibited significantly lower functional avidity than CD4+ T cells reactive to other viruses. Importantly, convalescents from mild COVID-19 had SARS-CoV-2 M-reactive CD4+ T cells with significantly lower functional avidity than convalescents from severe COVID-19. The current data suggest that pre-existing CCCoV M-specific memory CD4+ T cells may contribute to controlling SARS-CoV-2 infection by cross-reactivity, leading to mild disease but leaving memory cells with low functional avidity to SARS-CoV-2 M due to incomplete homology. These data provide indirect evidence that pre-existing cross-reactive CD4+ T cells contribute to protection from severe COVID-19.

Multimodal language models (MLMs) show promise for clinical decision support and diagnostic reasoning, raising the prospect of end-to-end automated medical image interpretation. However, clinicians are highly selective in adopting AI tools; a model that makes errors on seemingly simple perception tasks such as determining image orientation or identifying whether a CT scan is contrast-enhanced-are unlikely to be adopted for clinical tasks. We introduce MED-BLINK, a benchmark designed to probe these models for such perceptual abilities. MEDBLINK spans eight clinically meaningful tasks across multiple imaging modalities and anatomical regions, totaling 1,429 multiple-choice questions over 1,605 images. We evaluate 19 state-of-the-art MLMs, including general-purpose (GPT-4o, Claude 3.5 Sonnet) and domain-specific (Med-Flamingo, LLaVA-Med, RadFM) models. While human annotators achieve 96.4% accuracy, the best-performing model reaches only 65%. These results show that current MLMs frequently fail at routine perceptual checks, suggesting the need to strengthen their visual grounding to support clinical adoption. Data is available on our project page.

Background: During the 2023–2024 season, the influenza epidemic in South Korea peaked earlier, and the influenza vaccination rate among individuals aged ≥ 65 was high (82.2%). However, data on real-world vaccine effectiveness against influenza are lacking. Methods: From November 2023 to April 2024, we conducted a multicenter retrospective case–control study on adult patients aged ≥ 18 years who presented with influenza-like illness at seven medical centers as a part of a hospital-based influenza morbidity and mortality surveillance (HIMM) program in South Korea. Demographic and clinical data were collected from questionnaire surveys and electronic medical records. Using a test-negative design, we assessed the effectiveness of the 2023–2024 seasonal influenza vaccine, with age, sex, and comorbidities included as covariates. Results: A total of 3390 participants were enrolled through the HIMM system, including 1695 patients with either rapid antigen test (RAT) or real-time reverse-transcription polymerase chain reaction (RT-PCR) positive results and controls matched for age, sex, and months of registration. Among the 1696 influenza-positive patients, 1584 (93.5%) underwent RAT, with 88.9% testing positive for influenza A and 11.1% for influenza B. During the study periods, the overall vaccine effectiveness (VE) was 24.3% (95% confidence interval (CI), 11.5 to 35.2). The VE was insignificant when limited to older adults aged ≥ 65 years (13.5%; 95% CI, −17.9 to 36.6). In the subgroup analysis by subtype, the VE was 19.0% (95% CI, 5.0 to 31.0) for influenza A and 56.3% (95% CI, 35.3 to 70.6) for influenza B. Notably, influenza VE was 20.4% (95% CI, 2.9 to 34.8) in the early period (November to December) but decreased to 12.4% (95% CI, −14.9 to 33.2) in the late period (January to April). Conclusion: During the 2023–2024 season, the influenza vaccine showed a modest effectiveness (24.3%) against laboratory-confirmed influenza, which was particularly higher for influenza B. Because the VE was insignificant in older adults, particularly during the late period, better immunogenic influenza vaccines with longer-lasting protection should be considered.

Background: The World Health Organization has declared the end of the coronavirus disease 2019 (COVID-19) public health emergency. However, this did not indicate the end of COVID-19. Several months after the infection, numerous patients complain of respiratory or nonspecific symptoms; this condition is called long COVID. Even patients with mild COVID-19 can experience long COVID, thus the burden of long COVID remains considerable. Therefore, we conducted this study to comprehensively analyze the effects of long COVID using multi-faceted assessments. Materials and Methods: We conducted a prospective cohort study involving patients diagnosed with COVID-19 between February 2020 and September 2021 in six tertiary hospitals in Korea. Patients were followed up at 1, 3, 6, 12, 18, and 24 months after discharge. Long COVID was defined as the persistence of three or more COVID-19-related symptoms. The primary outcome of this study was the prevalence of long COVID after the period of COVID-19. Results: During the study period, 290 patients were enrolled. Among them, 54.5 and 34.6% experienced long COVID within 6 months and after more than 18 months, respectively. Several patients showed abnormal results when tested for post-traumatic stress disorder (17.4%) , anxiety (31.9%) after 18 months. In patients who underwent follow-up chest computed tomography 18 months after COVID-19, abnormal findings remained at 51.9%. Males (odds ratio [OR], 0.17; 95% confidence interval [CI], 0.05-0.53; P =0.004) and elderly (OR, 1.04; 95% CI, 1.00-1.09; P =0.04) showed a significant association with long COVID after 12-18 months in a multivariable logistic regression analysis. Conclusion: Many patients still showed long COVID after 18 months post SARS-CoV-2 infection. When managing these patients, the assessment of multiple aspects is necessary.

Background: Influenza cases have surged earlier than usual during the 2024–2025 season, with A/H1N1 (pdm09) being the dominant strain. We aimed to investigate early estimates of influenza vaccine effectiveness (VE) for the 2024–2025 season to enhance our influenza response strategies. Methods: From November 1 to December 31, 2024, we enrolled 990 individuals with influenza-like illness from the hospital-based influenza surveillance network (Hospital-Based Influenza Morbidity and Mortality, HIMM), which consists of eight hospitals. Results: The overall adjusted VE was estimated to be −0.5% (95% confidence interval [CI], −34.0 to 24.6), with 0.4% (95% CI, −33.2 to 25.5) for influenza A. Analyses by influenza subtype were exploratory, given the limited number of subtyped cases. Although ineffective in preventing laboratory-confirmed influenza, influenza vaccination reduced influenza-related hospitalizations by 31.9% (95% CI, 3.5 to 51.9). Conclusions: It is necessary to enhance influenza vaccine effectiveness by selecting better-matched vaccine strains and introducing immune-enhanced vaccines.

Health AI suffers from a systemic reproducibility crisis that irreparably hinders research in this space across academia and industry. To combat this and empower researchers in the health AI space, we propose a comprehensive interactive tutorial introducing the ''Medical Event Data Standard'' (MEDS) and its growing open-source ecosystem. Working in MEDS allows you to more easily build AI models over public or private longitudinal EHR datasets and to readily benchmark existing, published models against contributions on local datasets and tasks. MEDS simplifies the construction of AI models on longitudinal Electronic Health Record (EHR) datasets and enables straightforward benchmarking against established models. Reflecting its growing adoption, MEDS is utilized at over 15 institutions across 8 countries, features 7+ open-source tools, supports 10+ published models, and provides publicly available Extract-Transform-Load (ETL) pipelines for major public EHR datasets. A KDD tutorial offering practical experience with MEDS will significantly enhance reproducibility and comparability in health AI research. In this tutorial, we will teach attendees how to (1) transform datasets into the MEDS format(2) pre-process MEDS data for modeling needs(3) build highly effective, efficient, AI models for diverse predictive tasks on their datasets, and (4) contribute their results to MEDS-DEV, a decentralized benchmark enabling robust evaluation against meaningful baselines. Participants will engage in collaborative, minimal-dependency Jupyter notebook exercises, guided through each step by structured instruction and practical coding sessions. Attendees will leave equipped with practical knowledge to build reproducible, state-of-the-art AI models within the MEDS ecosystem.

Medical Decision-Making (MDM) is a multi-faceted process that requires clinicians to assess complex multi-modal patient data patient, often collaboratively. Large Language Models (LLMs) promise to streamline this process by synthesizing vast medical knowledge and multi-modal health data. However, single-agent are often ill-suited for nuanced medical contexts requiring adaptable, collaborative problem-solving. Our MDAgents addresses this need by dynamically assigning collaboration structures to LLMs based on task complexity, mimicking real-world clinical collaboration and decision-making. This framework improves diagnostic accuracy and supports adaptive responses in complex, real-world medical scenarios, making it a valuable tool for clinicians in various healthcare settings, and at the same time, being more efficient in terms of computing cost than static multi-agent decision making methods.

Severe fever with thrombocytopenia syndrome (SFTS) is an acute febrile illness caused by the SFTS virus (SFTSV). We conducted this study to propose a scientific evidence-based treatment that can improve prognosis through changes in viral load and inflammatory cytokines according to the specific treatment of SFTS patients. This prospective and observational study was conducted at 14 tertiary referral hospitals, which are located in SFTS endemic areas in Korea, from 1 May 2018 to 31 October 2020. Patients of any age were eligible for inclusion if they were polymerase chain reaction positive against SFTSV, or showed a four-fold or higher increase in IgG antibody titers between two serum samples collected during the acute and convalescent phases. On the other hand, patients with other tick-borne infections were excluded. In total, 79 patients were included in the study. The viral load of the group treated with steroids was 3.39, 3.21, and 1.36 log10 RNA copies/reaction at each week since the onset of symptoms, and the viral load in patients treated with plasma exchange was 4.47, 2.60, and 2.00 log10 RNA copies/reaction at each week after symptom onset. The inflammatory cytokines were not reduced effectively by any specific treatment except IVIG for the entire treatment period. Secondary infections according to pathogens revealed four bacterial (26.7%) and one fungal (6.7%) infection in the steroid group. The viral load of SFTSV and inflammatory cytokines cannot be decreased by steroid and plasma exchange treatments. Secondary bacterial infections can occur when steroids are administered for the treatment of SFTS. Therefore, caution should be exercised when choosing treatment strategies for SFTS.

In clinical practice, one often needs to identify whether a patient is at high risk of adverse outcomes after some key medical event. For example, quantifying the risk of adverse outcomes after an acute cardiovascular event helps healthcare providers identify those patients at the highest risk of poor outcomes; i.e., patients who benefit from invasive therapies that can lower their risk. Assessing the risk of adverse outcomes, however, is challenging due to the complexity, variability, and heterogeneity of longitudinal medical data, especially for individuals suffering from chronic diseases like heart failure. In this paper, we introduce Event-Based Contrastive Learning (EBCL) - a method for learning embeddings of heterogeneous patient data that preserves temporal information before and after key index events. We demonstrate that EBCL can be used to construct models that yield improved performance on important downstream tasks relative to other pretraining methods. We develop and test the method using a cohort of heart failure patients obtained from a large hospital network and the publicly available MIMIC-IV dataset consisting of patients in an intensive care unit at a large tertiary care center. On both cohorts, EBCL pretraining yields models that are performant with respect to a number of downstream tasks, including mortality, hospital readmission, and length of stay. In addition, unsupervised EBCL embeddings effectively cluster heart failure patients into subgroups with distinct outcomes, thereby providing information that helps identify new heart failure phenotypes. The contrastive framework around the index event can be adapted to a wide array of time-series datasets and provides information that can be used to guide personalized care.