{"paper_id":"41b66724-5f79-4aca-91ee-896e31f85aba","body_text":"Large Language Models in Infectious Diseases: A Systemic Review | Research Square window.SnipcartSettings = { analytics: { enabled: false } }; (function() { var accessVector = localStorage.getItem('access_vector') || ''; window.dataLayer = window.dataLayer || []; if (accessVector) { window.dataLayer.push({ user: { profile: { profileInfo: { snid: accessVector } } } }); } })(); (function(w,d,s,l,i){w[l]=w[l]||[];w[l].push({'gtm.start':new Date().getTime(),event:'gtm.js'});var f=d.getElementsByTagName(s)[0],j=d.createElement(s),dl=l!='dataLayer'?'&l='+l:'';j.async=true;j.src='https://www.googletagmanager.com/gtm.js?id='+i+dl;f.parentNode.insertBefore(j,f);})(window,document,'script','dataLayer','GTM-K279D39R'); Browse Preprints In Review Journals COVID-19 Preprints AJE Video Bytes Research Tools Research Promotion AJE Professional Editing AJE Rubriq About Preprint Platform In Review Editorial Policies Our Team Advisory Board Help Center Sign In Submit a Preprint Cite Share Download PDF Systematic Review Large Language Models in Infectious Diseases: A Systemic Review Alon Gorenshtein, Eyal Klang, Jacob J. Smith, Richard Dzeng, Mark C. Poznansky, and 2 more This is a preprint; it has not been peer reviewed by a journal. https://doi.org/ 10.21203/rs.3.rs-8901882/v1 This work is licensed under a CC BY 4.0 License Status: Posted Version 1 posted You are reading this latest preprint version Abstract Background Clinical reasoning in infectious diseases relies on validated evidence. LLMs are being introduced into diagnosis, antimicrobial stewardship, and guideline interpretation before their safety and reliability are established. Methods This review, registered in PROSPERO (CRD420251155354), evaluated studies using GPT, Claude, Gemini, and retrieval-augmented or agentic systems for infectious disease decision-making. PubMed, CENTRAL, Scopus, and Web of Science were searched from January 2018 to September 2025. Two reviewers screened and extracted data. Risk of bias was assessed with QUADAS-AI. Findings: Thirty-one studies met inclusion criteria. Most were cross-sectional (61%) and vignette-based (68%). Only 32% used real clinical data; 23% had low risk of bias. Safety issues were reported in 90% of studies: incomplete responses (61%), unsafe advice (23–32%), and fabricated content (32%). In antimicrobial stewardship, agreement with infectious-disease specialists was ~ 50%. Diagnostic sensitivity for structured infections was 80–100%. Retrieval-augmented systems increased specificity from 35% to 75% and reduced hallucinations. Proprietary models outperformed open-source models but did not reach expert accuracy. Interpretation: LLMs perform well in defined diagnostic tasks but remain unreliable for autonomous clinical use. High error rates, inconsistent reasoning, and fabricated content require expert oversight and external validation before deployment. Infectious Diseases Infectious diseases Large language models Clinical decision support Antimicrobial stewardship Hallucinations (AI) Patient safety Retrieval-augmented generation Bias and fairness Figures Figure 1 Figure 2 Figure 3 Main Point Summary Large language models show promise in infectious diseases for narrow diagnostic tasks, surveillance, and stewardship, but most evidence is retrospective and high risk of bias. Hallucinations and context errors persist. External validation, human oversight, and retrieval-augmented or agentic safeguards are essential. Introduction Clinical reasoning in infectious diseases (ID) depends on data, expertise, and judgment. Large language models (LLMs) now attempt to perform these tasks in text, generating diagnostic and therapeutic suggestions without task-specific training.¹ - ³ ID present a natural testing ground. The specialty combines complex decision-making, workforce shortages, and rising antimicrobial resistance (AMR).⁴ , ⁵ The clinical burden is high across settings, with infections accounting for roughly 10% of ambulatory encounters in the USA (higher in paediatric care) and healthcare-associated infections affecting about 3% of hospitalized patients in point-prevalence surveys, yet specialist access is uneven, as over 80% of counties lack an ID physician and many fellowship positions go unfilled.⁶ - ⁸ Against this backdrop global AMR caused an estimated 4.95 million deaths in 2019, concentrated in respiratory and bloodstream infections.⁹ This environment invites automation. LLMs have been proposed for surveillance, guideline interpretation, and epidemic prediction.¹⁰ - ¹⁵ Yet early reports show instability—hallucinations, unsafe recommendations, and vulnerability to manipulation.¹⁶ - ¹⁸ This review examines whether recent advances have turned LLMs into reliable instruments for ID care. Methods Protocol and reporting This systematic review was pre-registered with PROSPERO (CRD420251155354) and conducted according to PRISMA-2020 guidelines. 1 Although quantitative pooling was pre-specified in the protocol, substantial heterogeneity in populations, interventions, and outcomes precluded meta-analysis. Eligibility criteria We included primary research studies evaluating LLMs-including general-purpose models (e.g. ChatGPT, Gemini etc.), fine-tuned LLMs, retrieval-augmented generation (RAG) systems, and agentic LLMs-applied to ID decision-making in human clinical or public-health contexts. Eligible study designs were original research articles. Review papers, case reports, conference abstracts, editorials, preprints, and studies not conducted in English were excluded. Additionally, we excluded studies of classic machine learning or natural language processing methods without an LLM component, image-only models without LLM reasoning, non-ID applications, purely administrative or educational uses, and synthetic cases not traceable to real clinical scenarios. Studies focusing solely on user satisfaction without clinical correctness were also excluded. Information sources and search We searched PubMed, CENTRAL, Scopus, and Web of Science from 1 January 2018 to 30 September 2025. The search strategy combined index terms and free-text words for: (1) types of healthcare-associated infection and causative organisms; (2) LLM-based interventions, including IPC Core Components; (3) national or subnational context; and (4) outcome measures. No language restrictions were applied. Exact search strategies are provided in the appendix. Study selection Two independent reviewers screened titles (A.G., M.O.), abstracts, and full texts using pre-defined eligibility criteria. Disagreements were resolved by discussion or consultation with a third reviewer (E.K.). Data extraction Two reviewers (A.G. and M.O.) independently extracted key study details, including design, task type, disease focus, model used, data source, validation, comparator, and main outcomes. Unclear or missing information was clarified with authors or recorded as “not reported.” Risk of bias and reporting quality We assessed risk of bias using QUADAS-AI. 2 Two reviewers (A.G.,M.O.) independently assessed each domain; disagreements were resolved by discussion. Data synthesis We synthesized findings narratively, stratified by task category, disease focus, and LLM family. Within task categories, we grouped studies by specific infections (e.g., pneumonia/respiratory, bloodstream infection [BSI]/sepsis, catheter-associated urinary tract infection [CAUTI], invasive fungal infection [IFI/IMD], HIV/sexually transmitted infections [STIs], tuberculosis [TB]). When three or more commensurable studies examined the same task, disease, and metric, we reported descriptive statistics (median [IQR]) without formal meta-analytic pooling. Results Study selection The search retrieved 6,060 records from PubMed (1,329), CENTRAL (912), Scopus (3,379), and Web of Science (440). After removing 2,408 duplicates, 3,652 records were screened. Of these, 3,471 were excluded for irrelevance or pediatric focus. The remaining 181 full texts were reviewed, and 150 were excluded for not meeting inclusion criteria. Thirty-one studies, published between 2023 and 2025, were included in the final analysis ( Figure 1 ). 3–33 Risk of bias QUADAS-AI assessment revealed low overall risk in 7/31 studies (23%) and high risk in 24/31 (77%) ( Table Sup 1 ). Patient selection showed high risk in 24 studies (77%), primarily attributable to reliance on synthetic vignettes or curated guideline-based questions rather than consecutive real patient encounters. Meanwhile, the remaining domains exhibited predominantly low risk: index test (24/31 low risk, 77%), reference standard (27/31 low risk, 87%), and flow and timing (29/31 low risk, 94%). Study characteristics The 31 included studies evaluated LLM applications across diverse settings: inpatient wards (n=8), ICU (n=1), outpatient or primary care (n=1), mixed clinical settings (n=10), public health (n=2), laboratory (n=1), and comparative bench/question-based studies (n=8). Study designs comprised cross-sectional evaluations (n=19), prospective studies (n=5), retrospective analyses (n=4), and specialized accuracy/pilot studies (n=3). Disease foci included sexually transmitted diseases (n=4), urinary tract infections (n=4), tuberculosis (n=2), sepsis (n=2), HIV (n=2), hepatitis (n=2), mixed ID (n=3), and single studies on bacteremia, osteoarticular infections, vertebral osteomyelitis, blood culture diagnosis, surgical site infection, pneumonia, antimicrobial resistance, Infection Prevention and Control(IPC) guidelines, infective endocarditis prophylaxis, and general ID topics. Sample sizes ranged from 6 vignettes to 4,786 patients. LLM families included GPT-4/4o (n=20 studies), Claude 3/3.5 (n=7), Gemini (n=10), Mistral/Llama (n=3), custom chatbots (n=2), and specialized tools (n=6 studies using Consensus, OpenEvidence, Amboss, Perplexity, Microsoft Copilot). Retrieval-augmented generation or agentic architectures were explicitly implemented in 5 studies. Data sources were vignettes/synthetic cases (n=21) and real EHR or clinical data (n=10). Comparators included infectious disease specialists (n=8), guidelines (n=8), other LLMs (n=13), or none (n=2) ( Supplementary Table 2 ). Safety and Clinical Applicability Safety concerns were documented in 28/31 studies (90.3%). Clinically incomplete responses represented the most prevalent issue (19/31, 61.3%), with usefulness ratings of 42% and mean completeness scores of 40%. Unsafe recommendations occurred in 10/31 studies (32.3%), manifesting as substandard responses posing serious health risk in 10.3% of evaluated cases. Harmful or inadequate recommendations appeared in 7/31 studies (22.6%): 16% potentially hazardous bacteremia management plans (Maillard et al.) 32 , 71% incorrect isolation precautions in sepsis cases (Lorenzoni et al.) 3 , and 9% inadequate source-control recommendations rates significantly exceeding infectious disease expert benchmarks (1–4%, p<0.05). 32 Hallucinations (fluent, confident response that presents incorrect information, akin to clinical confabulation), fabricated citations, contradictory statements, or artificial clinical details were documented in 10/31 studies (32.3%). Constraint-induced limitations (guardrail refusals, unanswered queries, token limits preventing processing of complete medical records) affected 9/31 studies (29%): 8% of interactions yielded no response and 65% experienced extraction errors from incomplete chart access. Context-dependent failures (errors mitigated by providing complete clinical data or external knowledge bases) in 3 studies, (3/31, 9.7%) demonstrated marked improvement with interventions: retrieval-augmented generation reduced hallucinations; providing complete clinical documentation improved central line-associated bloodstream infection detection specificity from 35% to 75% (Rodriguez-Nava et al.). These findings indicate requirements for expert oversight, structured validation protocols, and context-enrichment architectures before clinical deployment (Supplementary Table 2). Antimicrobial Stewardship Performance Twelve studies assessed LLMs in antimicrobial stewardship. Across tasks, concordance with infectious-disease specialists was moderate, averaging about 50%. Two bacteremia vignette studies (n=100 each) reported identical 51% agreement (κ=0.48). Agreement was higher for Gram-positive (70%, κ=0.68) than Gram-negative infections (46%, κ=0.43). In Maillard et al.’s prospective study using real patient data (n=44, GPT-4), diagnostic accuracy reached 59% and empiric therapy appropriateness 64%, but 16% of recommendations were potentially harmful, including inactive agents and missed source-control interventions. 32 In a blood culture stewardship study (n = 84), LLMs produced 13% harmful or inadequate recommendations, significantly higher than experts (4%, p = 0.047). Most errors involved missing echocardiography for suspected endovascular infections. 24 A 14-model comparison showed wide variation in antibiotic prescribing accuracy. Antibiotic selection ranged from 30% to 100%, while dose and duration accuracy fell to 0–92%. Citation accuracy ranged 0–100%, and several models produced fabricated references. ChatGPT-o1 performed best overall (71.7% correct, 43/60). 30 In real-patient testing, Lorenzoni et al. (n = 7, GPT-4o) achieved perfect concordance for antibiotic selection but misjudged isolation precautions in 71% of cases. 3 Rodriguez-Nava and colleagues reported initially low central line-associated bloodstream infection (CLABSI) detection specificity (35%), which improved to 75% when complete chart information was provided, highlighting the critical role of RAG in improving accuracy. 11 In outpatient vignettes (n = 24, six models), Nguyen et al. found correct antibiotic selection between 59% and 100% and complete clinical advice between 25% and 96%, with proprietary models outperforming open-source ones. 29 Diagnostic Accuracy and Guideline Concordance Four studies evaluated LLM diagnostic accuracy for infectious conditions using sensitivity and specificity metrics. Diagnostic sensitivity ranged from 80% to 100% (median 91%), with high performance for catheter-associated urinary tract infection detection (91%), tuberculous pleural effusion diagnosis (89%), and surgical site infection screening (100%). Specificity varied substantially (range 35-100%, median 92%). Wu and colleagues demonstrated that a custom LLM for tuberculous pleural effusion diagnosis achieved AUROC 0·96 (sensitivity 76%, specificity 100%), matching or exceeding traditional machine learning models. 17 Satheakeerthy and colleagues showed zero-shot Llama-3-70B could screen surgical site infections with 100% sensitivity and 86% specificity, flagging infections earlier than infection control staff in 50% of cases. 12 Twelve studies evaluated LLMs in answering guideline-based questions, with accuracy ranging from 42% to 98% depending on topic and complexity. Borgonovo et al. found specialized RAG tools (Open Evidence, Microsoft Copilot) most accurate (94.4%), outperforming general-purpose models such as GPT-4o and Gemini 2.5 Pro (92.9%). 26 Lin et al. showed OpenAI O1 performed best for pneumonia guidelines, achieving 55% “excellent” responses and effective self-correction, compared with GPT-4o, which produced 40% “poor” responses. 20 Kufel et al. reported only 41.8% of GPT-3.5 outputs were rated “useful,” with low completeness (5.8/10) and safety (6.4/10). 28 Accuracy also varied by disease: tuberculosis questions scored 3.6–4.4/5, viral hepatitis 71–78%, and infective endocarditis prophylaxis 69–80%, with GPT-4o highest (80%). Performance was consistently better for informal, social-media-style questions (92%) than for formal guideline queries (69%, p < 0.001) ( Figure 2, Supplementary Table 3. ). 9 Limitations and Methodological Quality Methodological quality was constrained by design and reporting deficiencies. More than two-thirds of studies (21/31, 68%) evaluated LLMs using synthetic vignettes or curated cases, potentially overestimating performance by avoiding complexity, ambiguity, and incomplete documentation characteristic of actual practice. Among studies using real patient data (32%), most were retrospective with attendant selection bias. Sample sizes were frequently inadequate (35%), yielding unstable estimates. Validation approaches were weak across the evidence base. 12 studies (39%) had no expert comparison, instead benchmarking LLMs against other LLMs or guideline text alone. 14 studies (45%) judged correctness by guideline concordance without clinical context. The predominant cross-sectional design (22/31, 71%) precluded assessment of performance stability over time, while lack of blinding in 24/31 (77%) introduced expectation bias that might favor novel technology. Reproducibility and transparency were deficient. Most studies queried each question only once without assessing response variability, potentially concealing inconsistency and raising concerns about selective reporting of optimal responses. Process measures were absent in 21/31 (68%), leaving intervention fidelity and implementation strategies unclear. Outcome definitions were universally heterogeneous, employing non-standardized custom scales for usefulness, appropriateness, or quality that precluded cross-study comparison or meta-analysis ( Table 2 ). Discussion In this systematic review of 31 studies, we found wide variability in how LLMs perform across infectious disease medical practice. Their accuracy and reliability remain insufficient for autonomous clinical use. Across antimicrobial stewardship, diagnosis, guideline interpretation, and surveillance tasks, performance was inconsistent. Concordance with infectious-disease specialists for empiric therapy averaged about Taken together, the evidence suggests that LLM diagnostic performance is strongest for narrow, well-specified classification tasks, ranging from common syndromes such as CAUTI,SSI to rarer entities such as tuberculous pleural effusions. However, accuracy and calibration consistently decline when models must integrate context across comorbidities, timelines, devices, cultures, and competing diagnoses, which is where clinical reasoning is most vulnerable to error. According to current evidence, RAG can improve specificity by providing access to full clinical documentation. Yet safety remains a central limitation: most studies reported incomplete or unsafe recommendations and, in many cases, fabricated or contradictory content. Even the most advanced models (at time of evaluation), such as GPT-4o and Claude 3.5, performed better than open-source systems but still failed to reach expert reliability. Their stronger results on conversational or social-media-style questions compared with formal guideline queries suggest training data skewed toward lay information rather than specialist clinical knowledge. Safety and clinical applicability Safety has emerged as a significant barrier to the clinical deployment of LLMs in ID management. Rates of harmful or inadequate antimicrobial recommendations from these models ranged between 13% and 16%, sharply contrasting with the 1% to 4% error rates reported by infectious disease experts (Maillard et al.). Such harmful recommendations were described in Schwartz et al. document a paradigm case: when prompted to create a management plan for cryptococcal meningitis, GPT-3.5 recommended initiating antiretroviral therapy within 2 weeks-a recommendation 34 directly contradicted by Boulware et al.'s randomized controlled trial proving this approach increases mortality. 35 Such inaccuracies pose serious risks by jeopardizing patient safety and exacerbating the global issue of AMR. 36 The underlying causes of these errors may mirror those observed in AMR challenges globally, primarily arising from various factors, such as wrong indication, selection, dosage, duration, lack of adherence to infection prevention and control (IPC) protocols. 37,38 LLMs showed this issues when they were lacking a defined guidelines, leading to generate incorrect responses. This challenge is further compounded by geographic variability in treatment protocols (Nguyen et al.); as established LLMs often lack access to localized medical knowledge. 39 This highlights the imperative for implementing RAG systems, which would enable LLMs to integrate context-specific information, thereby aligning with established protocols and mitigating the risk of adverse treatments. Another contributing factor to these inaccuracies is the inherent issue of hallucinations generated by LLMs. 40 While not errors per se (byproduct of LLMs training) 41 , these hallucinations manifesting as spurious guideline citations, contradictory assertions within a single interaction, or fabricated clinical details were documented in approximately one-third of the studies reviewed, undermining clinical trust and introducing medico-legal risks. This phenomenon is linked to LLMs' limited access to real-time resources. 42 However, research indicates that providing web access significantly enhances their ability to generate accurate, high-quality scientific references. 43 Given these considerations, LLMs should not be deployed in their unmodified form due to their potential threats in the field ( Figure 3. ). Instead, they should be utilized as part of AI agents that leverage LLM capabilities while planning tasks, accessing external tools, and coordinating with other agents. In contrast to standard LLMs, these agents can perform multi-step processes, access real-time clinical information, and synthesize data from diverse sources. 44 This approach addresses the aforementioned safety concerns while also tackling additional issues such as verbosity, the need for expert-in-the-loop safety mechanisms, and iterative improvement. Due to the dearth of studies focusing specifically on ID, we cannot assertively conclude that such AI agents will resolve these challenges. However, this represents a critical area for future research as the next phase of LLM studies in infectious disease should aim to explore these innovations. CAUTI and CLABSI exemplify both the promise and the ceiling for LLM-enabled hospital epidemiology. These endpoints are operational quality metrics defined by NHSN surveillance rules, 45,46 not bedside diagnoses, so adjudication hinges on consistent application of standardized criteria to temporally ordered device, culture, and symptom data, often under substantial infection-prevention workload. 21 In CAUTI, GPT-4 achieved high performance on curated case abstractions and improved further with iterative, criteria-aligned prompting, highlighting how structured inputs and explicit rule framing can materially shift reliability. In contrast, CLABSI identification from real clinical notes was strongly context dependent: when the model was constrained to partial chart excerpts, sensitivity remained high but specificity was limited, and performance improved when key missing chart elements were supplied, supporting the use of RAG to pull the relevant EHR fields and NHSN rule elements before generation. Near term, LLMs are best deployed as definition-aware tools for education and structured abstraction that priorities sensitivity and trigger escalation, while final CAUTI and CLABSI attribution remains with trained infection-prevention reviewers. Because surveillance labels can propagate to antimicrobial decisions, CAUTI workflows should include explicit stewardship safeguards that confirm symptoms and exclude asymptomatic bacteriuria before outputs are acted on. Comparison with prior literature Up until now the topic of AI in ID is residing in a controversial place. A stark example is the contradiction between Siddig et al. declaring AI \"revolutionizing\" ID control 47 versus Schwartz's et al. \"Black Box Warning\" arguing \"existing LLMs are not safe for clinical consultation”. 34 Our own group’s previous systematic review of 15 studies identified promise of NLP and LLM in areas like pathogen detection and surveillance but noted limited real-world validation. 48 Our current findings extend beyond that review which included by the time of publications only two LLM studies. Be it as it may our current results show persistence of LLM limitation in literature, as evidenced by our observation that 58% of included studies relied on synthetic vignettes rather than real-world data. Howard et al. highlighted similar challenges in data completeness and interoperability for AI in tackling AMR, advocating for support of UN General Assembly targets like antimicrobial stewardship programs and surveillance, though with less focus on LLM-specific safety risks. 49 Despite this, our review found that LLMs may be useful in specific niches: diagnostics for urinary tract infections (UTIs), pneumonia, bloodstream infections (BSIs), and invasive fungal infections in defined populations (median AUROC 0.82, range 0.64–0.95); social media-based disease surveillance (accuracy 85–100%, with 92% for informal queries); antimicrobial stewardship (median appropriateness 71%, range 57–85%); and infection prevention/control with structured prompts (accuracy 98–100%). These successes share common features of well-defined diagnostic criteria, structured data sources, supplementary human verification, and consequences of errors that allow correction before patient harm. Limitations Our systematic review has several limitations. First, substantial heterogeneity in populations, interventions, and outcome measures precluded meta-analysis. Second, the overall risk of bias was high across included studies. Third, the rapidly evolving nature of LLM technology means newer models might show different performance characteristics than those evaluated in our included studies. Finally, the predominance of retrospective studies in our review indicates a need for more prospective studies to validate the findings. Conclusions The reviewed iterations of LLMs are not well-suited for clinical application in ID. Most studies highlight safety concerns despite the models demonstrating high performance in structured tasks (UTI, BSI and fungal infections). Issues such as hallucinations, missing guideline information, and lack of web search capabilities contribute to misinformation from LLMs. To mitigate these challenges, LLMs should be employed as AI agents before being utilized in future studies. Declarations Competing interests. The authors declare that they have no competing interests. Funding. This work was supported in part through the computational and data resources and staff expertise provided by Scientific Computing and Data at the Icahn School of Medicine at Mount Sinai and supported by the Clinical and Translational Science Awards (CTSA) grant UL1TR004419 from the National Center for Advancing Translational Sciences. Research reported in this publication was also supported by the Office of Research Infrastructure of the National Institutes of Health under award number S10OD026880 and S10OD030463. The content is solely the responsibility of the authors and does not necessarily represent the official views of the National Institutes of Health. Author Contributions: Conceptualization, AG, MO, EK, BSG, GN; Methodology, AG, MO, EK; Formal Analysis, AG, MO; Data Curation, AG, MO; Writing-Original Draft Preparation, AG, MO; Writing- Review & Editing, AG, MO, EK,MCP,RD,JJS, BSG,GN; Supervision: EK,GN References Page MJ, McKenzie JE, Bossuyt PM et al (2021) The PRISMA 2020 statement: an updated guideline for reporting systematic reviews. BMJ 372:n71. 10.1136/bmj.n71 Guni A, Sounderajah V, Whiting P, Bossuyt P, Darzi A, Ashrafian H (2024) Revised Tool for the Quality Assessment of Diagnostic Accuracy Studies Using AI (QUADAS-AI): Protocol for a Qualitative Study. JMIR Res Protoc 13:e58202. 10.2196/58202 Lorenzoni G, Garbin A, Brigiari G, Papappicco CAM, Manfrin V, Gregori D (2025) Large Language Models in Action: Supporting Clinical Evaluation in an Infectious Disease Unit. Healthcare 13(8):879. 10.3390/healthcare13080879 Evaluation of artificial intelligence (AI) chatbots for providing sexual health information: a consensus study using real-world clinical queries | BMC Public Health | Full Text. Accessed October 8 (2025) https://bmcpublichealth.biomedcentral.com/articles/ 10.1186/s12889-025-22933-8 De Vito A, Geremia N, Marino A et al (2025) Assessing ChatGPT’s theoretical knowledge and prescriptive accuracy in bacterial infections: a comparative study with infectious diseases residents and specialists. Infection 53(3):873–881. 10.1007/s15010-024-02350-6 AI Chatbots as Sources of STD Information: A Study on Reliability and Readability | Journal of Medical Systems. Accessed October 8 (2025) https://link.springer.com/article/ 10.1007/s10916-025-02178-z Zhuo KY, Kim P, Kovacic J et al (2024) Can Artificial Intelligence Treat My Urinary Tract Infections?—Evaluation of Health Information Provided by OpenAI ™ ChatGPT on Urinary Tract Infections. Société Int D’Urologie J 5(2):104–107. 10.3390/siuj5020018 Cakir H, Caglar U, Sekkeli S et al (2024) Evaluating ChatGPT ability to answer urinary tract Infection-Related questions. Infect Dis Now 54(4):104884. 10.1016/j.idnow.2024.104884 Rewthamrongsris P, Burapacheep J, Trachoo V, Porntaveetus T (2025) Accuracy of Large Language Models for Infective Endocarditis Prophylaxis in Dental Procedures. Int Dent J 75(1):206–212. 10.1016/j.identj.2024.09.033 Tunçer G, Güçlü KG (2024) How Reliable is ChatGPT as a Novel Consultant in Infectious Diseases and Clinical Microbiology? Infect Dis Clin Microbiol 6(1):55–59. 10.36519/idcm.2024.286 Rodriguez-Nava G, Egoryan G, Goodman KE, Morgan DJ, Salinas JL (2024) Performance of a large language model for identifying central line-associated bloodstream infections (CLABSI) using real clinical notes. Infect Control Hosp Epidemiol 46(3):1–4. 10.1017/ice.2024.164 Satheakeerthy S, Stretton B, Tsimiklis J et al Zero-shot large language model application for surgical site infection auditing. Infect Dis Health. Published online May 21, 2025:S2468-0451(25)00030 – 6. 10.1016/j.idh.2025.05.001 Cheah MH, Gan YN, Altice FL et al (2024) Testing the Feasibility and Acceptability of Using an Artificial Intelligence Chatbot to Promote HIV Testing and Pre-Exposure Prophylaxis in Malaysia: Mixed Methods Study. JMIR Hum Factors 11:e52055. 10.2196/52055 Ni Z, Oh S, Saifi R, Azwa I, Altice FL (2025) Evaluating the Usability of an HIV Prevention Artificial Intelligence Chatbot in Malaysia: National Observational Study. JMIR Hum Factors 12:e70034. 10.2196/70034 Li Y, Huang CK, Hu Y, Zhou XD, He C, Zhong JW (2025) Exploring the performance of large language models on hepatitis B infection-related questions: A comparative study. World J Gastroenterol 31(3):101092. 10.3748/wjg.v31.i3.101092 Comparison of the performances between ChatGPT and Gemini in answering questions on viral hepatitis | Scientific Reports. Accessed October 8 (2025) https://www.nature.com/articles/s41598-024-83575-1 The large language model diagnoses tuberculous pleural effusion in pleural effusion patients through clinical feature landscapes | Respiratory Research | Full Text. Accessed October 8 (2025) https://respiratory-research.biomedcentral.com/articles/ 10.1186/s12931-025-03130-y Large (2025) language models’ capabilities in responding to tuberculosis medical questions: testing ChatGPT, Gemini, and Copilot | Scientific Reports. Accessed October 8. https://www.nature.com/articles/s41598-025-03074-9 Giske CG, Bressan M, Fiechter F et al (2024) GPT-4-based AI agents-the new expert system for detection of antimicrobial resistance mechanisms? J Clin Microbiol 62(11):e0068924. 10.1128/jcm.00689-24 Lin Z, Li Y, Wu M et al (2025) Performance analysis of large language models Chatgpt-4o, OpenAI O1, and OpenAI O3 mini in clinical treatment of pneumonia: a comparative study. Clin Exp Med 25(1):213. 10.1007/s10238-025-01743-7 Wiemken TL, Carrico RM (2024) Assisting the infection preventionist: Use of artificial intelligence for health care-associated infection surveillance. Am J Infect Control 52(6):625–629. 10.1016/j.ajic.2024.02.007 Abosi OJ, Kobayashi T, Ross N et al (2024) A head-to-head comparison of the accuracy of commercially available large language models for infection prevention and control inquiries, 2024. Infect Control Hosp Epidemiol 46(3):1–3. 10.1017/ice.2024.205 Oh N, Cha WC, Seo JH et al (2024) ChatGPT Predicts In-Hospital All-Cause Mortality for Sepsis: In-Context Learning with the Korean Sepsis Alliance Database. Healthc Inf Res 30(3):266–276. 10.4258/hir.2024.30.3.266 Tassone DM, Hitchcock MM, Rossier CJ et al (2025) Evaluating chain-of-thought prompting in a GPT chatbot for BCID2 interpretation and stewardship: how does AI compare to human experts? Antimicrob Steward Healthc Epidemiol ASHE 5(1):e154. 10.1017/ash.2025.10059 Milicevic F, Ghandour M, Khasawneh MY et al (2025) Assessing LLMs on IDSA Practice Guidelines for the Diagnosis and Treatment of Native Vertebral Osteomyelitis: A Comparison Study. J Clin Med 14(14):4996. 10.3390/jcm14144996 Borgonovo F, Matsuo T, Petri F et al (2025) Battle of the Bots: Solving Clinical Cases in Osteoarticular Infections With Large Language Models. Mayo Clin Proc Digit Health 3(3):100230. 10.1016/j.mcpdig.2025.100230 Evaluating the Performance of State-of-the-Art Artificial Intelligence Chatbots Based on the WHO Global Guidelines for the Prevention of Surgical Site Infection: Cross-Sectional Study - PubMed. Accessed October 8 (2025) https://pubmed.ncbi.nlm.nih.gov/40744114/ Kufel WD, Hanrahan KD, Seabury RW et al (2024) Let’s Have a Chat: How Well Does an Artificial Intelligence Chatbot Answer Clinical Infectious Diseases Pharmacotherapy Questions? Open Forum Infect Dis 11(11):ofae641. 10.1093/ofid/ofae641 Ngoc Nguyen O, Amin D, Bennett J et al (2025) GP or ChatGPT? Ability of large language models (LLMs) to support general practitioners when prescribing antibiotics. J Antimicrob Chemother 80(5):1324–1330. 10.1093/jac/dkaf077 De Vito A, Geremia N, Bavaro DF et al (2025) Comparing large language models for antibiotic prescribing in different clinical scenarios: which performs better? Clin Microbiol Infect Off Publ Eur Soc Clin Microbiol Infect Dis 31(8):1336–1342. 10.1016/j.cmi.2025.03.002 Montiel-Romero S, Rajme-López S, Román-Montes CM et al (2025) Recommended antibiotic treatment agreement between infectious diseases specialists and ChatGPT®. BMC Infect Dis 25(1):38. 10.1186/s12879-024-10426-9 Maillard A, Micheli G, Lefevre L et al (2024) Can Chatbot Artificial Intelligence Replace Infectious Diseases Physicians in the Management of Bloodstream Infections? A Prospective Cohort Study. Clin Infect Dis Off Publ Infect Dis Soc Am 78(4):825–832. 10.1093/cid/ciad632 Perret J, Schmid A (2024) Application of OpenAI GPT-4 for the retrospective detection of catheter-associated urinary tract infections in a fictitious and curated patient data set. Infect Control Hosp Epidemiol 45(1):96–99. 10.1017/ice.2023.189 Schwartz IS, Link KE, Daneshjou R, Cortés-Penfield N (2023) Black Box Warning: Large Language Models and the Future of Infectious Diseases Consultation. Clin Infect Dis Off Publ Infect Dis Soc Am 78(4):860–866. 10.1093/cid/ciad633 Timing of Antiretroviral Therapy after Diagnosis of Cryptococcal Meningitis | New England Journal of Medicine. Accessed October 7 (2025) https://www.nejm.org/doi/full/ 10.1056/NEJMoa1312884 GBD 2021 Antimicrobial Resistance Collaborators (2024) Global burden of bacterial antimicrobial resistance 1990–2021: a systematic analysis with forecasts to 2050. Lancet Lond Engl 404(10459):1199–1226. 10.1016/S0140-6736(24)01867-1 Otaigbe II, Elikwu CJ (2023) Drivers of inappropriate antibiotic use in low- and middle-income countries. JAC-Antimicrob Resist 5(3):dlad062. 10.1093/jacamr/dlad062 Mulchandani R, Tiseo K, Nandi A et al (2025) Global trends in inappropriate use of antibiotics, 2000–2021: scoping review and prevalence estimates. BMJ Public Health 3(1):e002411. 10.1136/bmjph-2024-002411 Busch F, Hoffmann L, Rueger C et al (2025) Current applications and challenges in large language models for patient care: a systematic review. Commun Med 5(1):1–13. 10.1038/s43856-024-00717-2 Omar M, Sorin V, Collins JD et al (2025) Multi-model assurance analysis showing large language models are highly vulnerable to adversarial hallucination attacks during clinical decision support. Commun Med 5(1):330. 10.1038/s43856-025-01021-3 Kalai AT, Nachum O, Vempala SS, Zhang E (2025) Why Language Models Hallucinate. arXiv . Preprint posted online September 4. 10.48550/arXiv.2509.04664 Generating credible referenced medical research: A comparative study of openAI’s GPT-4 and Google’s gemini - ScienceDirect. Accessed December 16 (2024) https://www.sciencedirect.com/science/article/pii/S0010482524016305 Gorenshtein A, Shihada K, Sorka M, Aran D, Shelly S (2025) LITERAS: Biomedical literature review and citation retrieval agents. Comput Biol Med 192:110363. 10.1016/j.compbiomed.2025.110363 Gorenshtein A, Omar M, Glicksberg BS, Nadkarni GN, Klang E AI Agents in Clinical Medicine: A Systematic Review. medRxiv . Preprint posted online August 26, 2025:2025.08.22.25334232. 10.1101/2025.08.22.25334232 Bloodstream Infections Published online 2025 Urinary Tract Infection Published online 2025 Siddig EE, Eltigani HF, Ahmed A (2023) The Rise of AI: How Artificial Intelligence is Revolutionizing Infectious Disease Control. Ann Biomed Eng 51(12):2636–2637. 10.1007/s10439-023-03280-4 Omar M, Brin D, Glicksberg B, Klang E (2024) Utilizing natural language processing and large language models in the diagnosis and prediction of infectious diseases: A systematic review. Am J Infect Control 52(9):992–1001. 10.1016/j.ajic.2024.03.016 Howard A, Reza N, Green PL et al (2025) Artificial intelligence and infectious diseases: tackling antimicrobial resistance, from personalised care to antibiotic discovery. Lancet Infect Dis 0(0). 10.1016/S1473-3099(25)00313-5 Tables Table 1. Safety issue categories, prevalence across studies, quantitative signals, and mitigations Safety issue category Studies (n/31; %) Representative quantitative rate(s) Severity tier Context-dependent? Clinical risk domain(s) Common mitigations Example studies Unsafe recommendations 10/31 (32.3%) 7/68 (10.3%) substandard responses posing serious health risk High Yes Guideline adherence, dosing, antimicrobial choice, AMR interpretation Expert review, external validation Abosi 2024, Cakir 2023 Harmful/inadequate recommendations 7/31 (22.6%) 7/44 (15.9%) harmful management plans; 5/7 (71.4%) incorrect isolation precautions; 4/44 (9.1%) inadequate source control High Yes Isolation/IPC, antimicrobial choice, dosing, source control Prompt engineering, chain-of-thought prompting, expert review Lorenzoni 2025, Maillard 2024 Hallucinations / fabricated citations 10/31 (32.3%) not reported Moderate Yes Antimicrobial choice, dosing, citations RAG to guidelines, external knowledge bases Borgonovo 2025, Cakir 2023 Inadequate or incomplete responses Clinically incomplete 19/31 (61.3%) 209/500 (41.8%) responses deemed not useful; 8/20 (40.0%) poor guideline adherence; 42/100 (42.0%) AMR mechanism errors Moderate Yes Isolation/IPC, guideline adherence, antimicrobial choice, citations, AMR interpretation, diagnostics, dosing Chain-of-thought prompting, prompt engineering, external validation Abosi 2024, Kufel 2024 Constraint-induced 9/31 (29.0%) 31/393 (7.9%) unanswered queries; 11/17 (64.7%) errors from missing chart information Low Yes Diagnostics, extraction RAG, expanded context windows, structured input templates Lorenzoni 2025, Rodriguez-Nava 2025 Internal contradictions / instability 2/31 (6.5%) not reported Moderate Unclear Guideline adherence, antimicrobial choice Version pinning, multiple runs with consensus Cakir 2023, Borgonovo 2025 Context-dependent failures 3/31 (9.7%) Specificity improved from 35% to 75% when complete chart provided Moderate Yes Diagnostics, surveillance RAG, full chart access, external data integration Rodriguez-Nava 2025, Wiemken 2024 Any safety event 28/31 (90.3%) — — — — — — Footnotes: Definitions: Unsafe recommendations = clearly incorrect/contraindicated management; Harmful/inadequate = omissions (e.g., source control) or dosing errors with potential for patient harm; Hallucinations = fabricated citations, contradictory statements, or artificial clinical details; Clinically incomplete = missing key clinical elements despite being prompted; Constraint-induced = refusals, guardrail blocks, verbosity, token limits, or language barriers preventing complete responses; Internal contradictions = intra-session inconsistencies in recommendations; Context-dependent = errors mitigated by providing full chart access or external knowledge. Percentages reflect unique studies per category. Studies may appear in multiple categories if they reported multiple safety issue types. \"Not reported\" indicates no numeric denominator was available in the source studies despite qualitative safety signals being documented. \"Context-dependent\" indicates issues that improved with RAG, full chart access, or expert oversight in reported studies. Common mitigations listed represent strategies that were tested and showed some benefit across studies, though not all were formally quantified. Table 2. Study Limitations and Clinical Applicability Constraints Limitation Category Studies Affected (n/31; %) Description and Examples Impact on Clinical Applicability Study Design and Validation Vignette/synthetic cases 21/31 (67.74%) Simulated rather than real clinical scenarios; may not capture complexity May overestimate performance in actual practice Small sample size 11/31 (35%) <50 cases evaluated (range: 7-44) Insufficient power; wide confidence intervals Single-center studies 8/31 (26%) Limited to one institution's practices/population Generalizability uncertain across healthcare systems Data and Context Constraints Language limitations 3/31 (10%) Clinical documentation in non-English (Italian, Turkish); chatbots English-only Reduced accuracy for non-English contexts; limits global applicability Token/context limits 5/31 (16%) Unable to process complete medical records; limited to recent 2 progress notes Critical information missed; specificity dropped 35%→75% when full chart provided Knowledge cutoff issues 4/31 (13%) Training data ended 2021; inconsistent with latest guidelines (e.g., 2023 HBV consensus) Outdated recommendations; missed recent evidence Missing local context 6/31 (19%) No integration of local antibiograms, formularies, or resistance patterns Recommendations may be inappropriate/unavailable regionally Governance and Safety No PHI protections 12/31 (39%) Public web interfaces; no HIPAA-compliant/secure instances Cannot be used with real patient data in most settings IRB not required/waived 8/31 (26%) Classified as non-human subjects research or quality improvement Ethical oversight gaps for clinical deployment Lack of regulatory alignment 23/31 (74%) No FDA/CE marking or clinical validation pathway Unclear regulatory status for clinical use Implementation and Usability Verbosity/poor readability 5/31 (16%) Responses 200-400+ words; reading level >10th grade vs recommended 6th grade Unusable in time-critical settings; incomprehensible to patients Inconsistent reproducibility 4/31 (13%) Different answers when asked repeatedly; within-session contradictions Unreliable for clinical decision-making Constraint-induced failures 9/31 (29%) Guardrail refusals, unanswered queries (8% of interactions), or excessive caveats Incomplete clinical utility; frustrating user experience No uncertainty quantification 6/31 (19%) Failed to acknowledge limitations or express appropriate uncertainty False confidence; safety risk Clinical Preparedness Gaps No real-time EHR integration 27/31 (87%) Manual data entry required; cannot access lab/imaging/notes automatically Impractical for clinical workflow; introduces transcription errors No human-in-the-loop validation 19/31 (61%) No expert review mechanism before recommendations acted upon Unsafe for autonomous use Insufficient process measures 21/31 (68%) Intervention fidelity not reported; unclear what drove observed effects Cannot determine which implementation strategies effective Footnotes: Percentages represent unique studies reporting each limitation category. Categories are not mutually exclusive. EHR=electronic health record; HBV=hepatitis B virus; IRB=institutional review board; PHI=protected health information. Additional Declarations The authors declare no competing interests. Supplementary Files FullAPPENDIXIDFinalFeb.pdf Full Appendix Cite Share Download PDF Status: Posted Version 1 posted You are reading this latest preprint version Research Square lets you share your work early, gain feedback from the community, and start making changes to your manuscript prior to peer review in a journal. As a division of Research Square Company, we’re committed to making research communication faster, fairer, and more useful. We do this by developing innovative software and high quality services for the global research community. Our growing team is made up of researchers and industry professionals working together to solve the most critical problems facing scientific publishing. Also discoverable on Platform About Our Team In Review Editorial Policies Advisory Board Help Center Resources Author Services Accessibility API Access RSS feed Manage Cookie Preferences © Research Square 2026 | ISSN 2693-5015 (online) Privacy Policy Terms of Service Do Not Sell My Personal Information {\"props\":{\"pageProps\":{\"initialData\":{\"identity\":\"rs-8901882\",\"acceptedTermsAndConditions\":true,\"allowDirectSubmit\":true,\"archivedVersions\":[],\"articleType\":\"Systematic Review\",\"associatedPublications\":[],\"authors\":[{\"id\":592798085,\"identity\":\"f4341927-e38a-43d9-b590-0ecf40ace072\",\"order_by\":0,\"name\":\"Alon Gorenshtein\",\"email\":\"data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAAZAAAAAyAQMAAABI0h/eAAAABlBMVEX///8AAABVwtN+AAAACXBIWXMAAA7EAAAOxAGVKw4bAAABC0lEQVRIiWNgGAWjYBCDBH4JxgYJEAPCZwPiAwS0SM4gWYvBDQYG4rTw9y8+9uHjjro849vNjbdu1DDk8c8+Y/i5ooxBju9GAlYtEjeeJc+ceeZwsdmdg83WOccYiiXO5RhLnjnHYCyJQwvDjTPGzLxtBxK33Uhsk85tYEhsOMOWINnYxpC4AYcWeZCWv211iZtnQLXMP8OW/BOopR6XFoPzPcbMjG3MiRskoFo2nGE+BrIFGCDYtRjeYEtm7G07nDgD4heJxI1ALZYN5yQMZ555gFWL3PnDhxl+Ah3WP7v94e2cGpvEeWcYm282lNnI8x3H4X0JVHEJDAYm4D+AW24UjIJRMApGARgAACqEahXT3MrWAAAAAElFTkSuQmCC\",\"orcid\":\"https://orcid.org/0009-0000-7542-8608\",\"institution\":\"Department of Neurology, Harvard Medical School, Boston, MA\",\"correspondingAuthor\":true,\"prefix\":\"\",\"firstName\":\"Alon\",\"middleName\":\"\",\"lastName\":\"Gorenshtein\",\"suffix\":\"\"},{\"id\":592798086,\"identity\":\"80424d75-ca72-4c3c-84f1-bab9a8047194\",\"order_by\":1,\"name\":\"Eyal Klang\",\"email\":\"\",\"orcid\":\"https://orcid.org/0000-0002-4567-3108\",\"institution\":\"The Windreich Department of Artificial Intelligence and Human Health, Mount Sinai Medical Center, NY, USA.\",\"correspondingAuthor\":false,\"prefix\":\"\",\"firstName\":\"Eyal\",\"middleName\":\"\",\"lastName\":\"Klang\",\"suffix\":\"\"},{\"id\":592798087,\"identity\":\"cf94946a-afd5-459c-9606-1646bd4ff545\",\"order_by\":2,\"name\":\"Jacob J. Smith\",\"email\":\"\",\"orcid\":\"\",\"institution\":\"Vaccine and Immunotherapy Center, Massachusetts General Hospital, Boston, MA 02129, USA\",\"correspondingAuthor\":false,\"prefix\":\"\",\"firstName\":\"Jacob\",\"middleName\":\"J.\",\"lastName\":\"Smith\",\"suffix\":\"\"},{\"id\":592798088,\"identity\":\"89bedfa7-5317-4ee8-9afe-d42c5f250702\",\"order_by\":3,\"name\":\"Richard Dzeng\",\"email\":\"\",\"orcid\":\"\",\"institution\":\"Vaccine and Immunotherapy Center, Massachusetts General Hospital, Boston, MA 02129, USA\",\"correspondingAuthor\":false,\"prefix\":\"\",\"firstName\":\"Richard\",\"middleName\":\"\",\"lastName\":\"Dzeng\",\"suffix\":\"\"},{\"id\":592798089,\"identity\":\"3a8012f1-793a-4271-b004-05e96b48929a\",\"order_by\":4,\"name\":\"Mark C. Poznansky\",\"email\":\"\",\"orcid\":\"\",\"institution\":\"Vaccine and Immunotherapy Center, Massachusetts General Hospital, Boston, MA 02129, USA\",\"correspondingAuthor\":false,\"prefix\":\"\",\"firstName\":\"Mark\",\"middleName\":\"C.\",\"lastName\":\"Poznansky\",\"suffix\":\"\"},{\"id\":592798090,\"identity\":\"363a7592-957d-44cf-881d-04256a36809a\",\"order_by\":5,\"name\":\"Girish N Nadkarni\",\"email\":\"\",\"orcid\":\"https://orcid.org/0000-0001-6319-4314\",\"institution\":\"The Windreich Department of Artificial Intelligence and Human Health, Mount Sinai Medical Center, NY, USA.\",\"correspondingAuthor\":false,\"prefix\":\"\",\"firstName\":\"Girish\",\"middleName\":\"N\",\"lastName\":\"Nadkarni\",\"suffix\":\"\"},{\"id\":592798091,\"identity\":\"7bcc817d-4db3-4ff2-81e7-5bb46503b1d8\",\"order_by\":6,\"name\":\"Mahmud Omar\",\"email\":\"\",\"orcid\":\"https://orcid.org/0009-0001-0438-0827\",\"institution\":\"The Windreich Department of Artificial Intelligence and Human Health, Mount Sinai Medical Center, NY, USA.\",\"correspondingAuthor\":false,\"prefix\":\"\",\"firstName\":\"Mahmud\",\"middleName\":\"\",\"lastName\":\"Omar\",\"suffix\":\"\"}],\"badges\":[],\"createdAt\":\"2026-02-17 14:10:59\",\"currentVersionCode\":1,\"declarations\":{\"humanSubjects\":true,\"vertebrateSubjects\":false,\"conflictsOfInterestStatement\":false,\"humanSubjectEthicalGuidelines\":true,\"humanSubjectConsent\":true,\"humanSubjectClinicalTrial\":false,\"humanSubjectCaseReport\":false,\"vertebrateSubjectEthicalGuidelines\":false},\"doi\":\"10.21203/rs.3.rs-8901882/v1\",\"doiUrl\":\"https://doi.org/10.21203/rs.3.rs-8901882/v1\",\"draftVersion\":[],\"editorialEvents\":[],\"editorialNote\":\"\",\"failedWorkflow\":false,\"files\":[{\"id\":102898494,\"identity\":\"a891ac53-e6a5-40f4-b635-74dd375820b7\",\"added_by\":\"auto\",\"created_at\":\"2026-02-18 07:27:58\",\"extension\":\"png\",\"order_by\":1,\"title\":\"Figure 1\",\"display\":\"\",\"copyAsset\":false,\"role\":\"figure\",\"size\":159139,\"visible\":true,\"origin\":\"\",\"legend\":\"\\u003cp\\u003e\\u003cstrong\\u003ePRISMA Flow Diagram\\u003c/strong\\u003e\\u003c/p\\u003e\",\"description\":\"\",\"filename\":\"Figure1.png\",\"url\":\"https://assets-eu.researchsquare.com/files/rs-8901882/v1/126287ea283379884a5caa78.png\"},{\"id\":102898495,\"identity\":\"8a28c137-fdae-4eb2-99c8-e404fa7daefc\",\"added_by\":\"auto\",\"created_at\":\"2026-02-18 07:27:58\",\"extension\":\"png\",\"order_by\":2,\"title\":\"Figure 2\",\"display\":\"\",\"copyAsset\":false,\"role\":\"figure\",\"size\":141531,\"visible\":true,\"origin\":\"\",\"legend\":\"\\u003cp\\u003e\\u003cstrong\\u003ePerformance of large language models across clinical task categories in infectious diseases: \\u003c/strong\\u003ePerformance metrics reported across 30 studies evaluating large language models in infectious diseases, stratified by clinical task category. Each point represents an individual performance metric from a single study, with colour indicating metric type: accuracy (blue), sensitivity (green), specificity (orange), concordance (purple), or AUROC (grey). White diamonds indicate median performance within each task category. Vertical lines represent the range from minimum to maximum reported values. Numbers in parentheses denote the number of studies contributing to each category. Performance varied substantially within and across task categories, with antimicrobial stewardship and guideline-based question-answering showing the widest ranges (0–100% and 21–93%, respectively), while diagnosis and infection prevention tasks demonstrated more consistently high performance (medians \\u0026gt;90%). The single prognosis study reported AUROC of 66% for sepsis mortality prediction, marginally above chance level (50%, indicated).\\u003c/p\\u003e\",\"description\":\"\",\"filename\":\"Figure2.png\",\"url\":\"https://assets-eu.researchsquare.com/files/rs-8901882/v1/e800b7f58574cf41b7bcc419.png\"},{\"id\":102898496,\"identity\":\"962bb998-b3d9-4e9b-b8ad-d0937655735d\",\"added_by\":\"auto\",\"created_at\":\"2026-02-18 07:27:58\",\"extension\":\"png\",\"order_by\":3,\"title\":\"Figure 3\",\"display\":\"\",\"copyAsset\":false,\"role\":\"figure\",\"size\":106589,\"visible\":true,\"origin\":\"\",\"legend\":\"\\u003cp\\u003e\\u003cstrong\\u003eThreats of Large Language Models in Infectious Diseases: \\u003c/strong\\u003eInaccurate or context-blind outputs can drive inappropriate treatment, antimicrobial resistance, longer hospital stays and costs, medico-legal exposure, and outbreak escalation, showcasing the need for expert oversight, validated RAG, and regulation before clinical deployment.\\u003c/p\\u003e\",\"description\":\"\",\"filename\":\"Figure3.png\",\"url\":\"https://assets-eu.researchsquare.com/files/rs-8901882/v1/23f94df81d7796d2c1fd4ebe.png\"},{\"id\":102964204,\"identity\":\"b6443ca8-0417-4c62-8a51-1e328446ce33\",\"added_by\":\"auto\",\"created_at\":\"2026-02-19 04:21:43\",\"extension\":\"pdf\",\"order_by\":0,\"title\":\"\",\"display\":\"\",\"copyAsset\":false,\"role\":\"manuscript-pdf\",\"size\":1276149,\"visible\":true,\"origin\":\"\",\"legend\":\"\",\"description\":\"\",\"filename\":\"manuscript.pdf\",\"url\":\"https://assets-eu.researchsquare.com/files/rs-8901882/v1/c382efee-0f8a-4da5-941c-d0534affffb3.pdf\"},{\"id\":102898497,\"identity\":\"4b2d4e6e-a7f4-42a5-9894-39f34490a5bf\",\"added_by\":\"auto\",\"created_at\":\"2026-02-18 07:27:58\",\"extension\":\"pdf\",\"order_by\":1,\"title\":\"\",\"display\":\"\",\"copyAsset\":false,\"role\":\"supplement\",\"size\":1246612,\"visible\":true,\"origin\":\"\",\"legend\":\"\\u003cp\\u003eFull Appendix\\u003c/p\\u003e\",\"description\":\"\",\"filename\":\"FullAPPENDIXIDFinalFeb.pdf\",\"url\":\"https://assets-eu.researchsquare.com/files/rs-8901882/v1/8ec5efd31b6966c52d5b409b.pdf\"}],\"financialInterests\":\"The authors declare no competing interests.\",\"formattedTitle\":\"\\u003cp\\u003e\\u003cstrong\\u003eLarge Language Models in Infectious Diseases: A Systemic Review\\u003c/strong\\u003e\\u003c/p\\u003e\",\"fulltext\":[{\"header\":\"Main Point Summary\",\"content\":\"\\u003cp\\u003eLarge language models show promise in infectious diseases for narrow diagnostic tasks, surveillance, and stewardship, but most evidence is retrospective and high risk of bias. Hallucinations and context errors persist. External validation, human oversight, and retrieval-augmented or agentic safeguards are essential.\\u003c/p\\u003e\"},{\"header\":\"Introduction\",\"content\":\"\\u003cp\\u003eClinical reasoning in infectious diseases (ID) depends on data, expertise, and judgment. Large language models (LLMs) now attempt to perform these tasks in text, generating diagnostic and therapeutic suggestions without task-specific training.\\u0026sup1;\\u003csup\\u003e-\\u003c/sup\\u003e\\u0026sup3;\\u003c/p\\u003e\\n\\u003cp\\u003eID present a natural testing ground. The specialty combines complex decision-making, workforce shortages, and rising antimicrobial resistance (AMR).⁴\\u003csup\\u003e,\\u003c/sup\\u003e⁵ The clinical burden is high across settings, with infections accounting for roughly 10% of ambulatory encounters in the USA (higher in paediatric care) and healthcare-associated infections affecting about 3% of hospitalized patients in point-prevalence surveys, yet specialist access is uneven, as over 80% of counties lack an ID physician and many fellowship positions go unfilled.⁶\\u003csup\\u003e-\\u003c/sup\\u003e⁸ Against this backdrop global AMR caused an estimated 4.95 million deaths in 2019, concentrated in respiratory and bloodstream infections.⁹\\u003c/p\\u003e\\n\\u003cp\\u003eThis environment invites automation. LLMs have been proposed for surveillance, guideline interpretation, and epidemic prediction.\\u0026sup1;⁰\\u003csup\\u003e-\\u003c/sup\\u003e\\u0026sup1;⁵ Yet early reports show instability\\u0026mdash;hallucinations, unsafe recommendations, and vulnerability to manipulation.\\u0026sup1;⁶\\u003csup\\u003e-\\u003c/sup\\u003e\\u0026sup1;⁸\\u003c/p\\u003e\\n\\u003cp\\u003eThis review examines whether recent advances have turned LLMs into reliable instruments for ID care.\\u003c/p\\u003e\"},{\"header\":\"Methods\",\"content\":\"\\u003cp\\u003e\\u003cem\\u003eProtocol and reporting\\u003c/em\\u003e\\u003c/p\\u003e\\n\\u003cp\\u003eThis systematic review was pre-registered with PROSPERO (CRD420251155354) and conducted according to PRISMA-2020 guidelines.\\u003csup\\u003e1\\u003c/sup\\u003e Although quantitative pooling was pre-specified in the protocol, substantial heterogeneity in populations, interventions, and outcomes precluded meta-analysis.\\u003c/p\\u003e\\n\\u003cp\\u003e\\u003cem\\u003eEligibility criteria\\u003c/em\\u003e\\u003c/p\\u003e\\n\\u003cp\\u003eWe included primary research studies evaluating LLMs-including general-purpose models (e.g. ChatGPT, Gemini etc.), fine-tuned LLMs, retrieval-augmented generation (RAG) systems, and agentic LLMs-applied to ID decision-making in human clinical or public-health contexts.\\u0026nbsp;\\u003c/p\\u003e\\n\\u003cp\\u003eEligible study designs were original research articles. Review papers, case reports, conference abstracts, editorials, preprints, and studies not conducted in English were excluded. Additionally, we excluded studies of classic machine learning or natural language processing methods without an LLM component, image-only models without LLM reasoning, non-ID applications, purely administrative or educational uses, and synthetic cases not traceable to real clinical scenarios. Studies focusing solely on user satisfaction without clinical correctness were also excluded.\\u0026nbsp;\\u003c/p\\u003e\\n\\u003cp\\u003e\\u003cem\\u003eInformation sources and search\\u003c/em\\u003e\\u003c/p\\u003e\\n\\u003cp\\u003eWe searched PubMed, CENTRAL, Scopus, and Web of Science from 1 January 2018 to 30 September 2025. The search strategy combined index terms and free-text words for: (1) types of healthcare-associated infection and causative organisms; (2) LLM-based interventions, including IPC Core Components; (3) national or subnational context; and (4) outcome measures. No language restrictions were applied. Exact search strategies are provided in the appendix.\\u003c/p\\u003e\\n\\u003cp\\u003e\\u003cem\\u003eStudy selection\\u003c/em\\u003e\\u003c/p\\u003e\\n\\u003cp\\u003eTwo independent reviewers screened titles (A.G., M.O.), abstracts, and full texts using pre-defined eligibility criteria. Disagreements were resolved by discussion or consultation with a third reviewer (E.K.). \\u0026nbsp;\\u003c/p\\u003e\\n\\u003cp\\u003e\\u003cem\\u003eData extraction\\u003c/em\\u003e\\u003c/p\\u003e\\n\\u003cp\\u003eTwo reviewers (A.G. and M.O.) independently extracted key study details, including design, task type, disease focus, model used, data source, validation, comparator, and main outcomes. Unclear or missing information was clarified with authors or recorded as \\u0026ldquo;not reported.\\u0026rdquo;\\u003c/p\\u003e\\n\\u003cp\\u003e\\u0026nbsp;\\u003c/p\\u003e\\n\\u003cp\\u003e\\u003cem\\u003eRisk of bias and reporting quality\\u003c/em\\u003e\\u003c/p\\u003e\\n\\u003cp\\u003eWe assessed risk of bias using QUADAS-AI.\\u003csup\\u003e2\\u003c/sup\\u003e Two reviewers (A.G.,M.O.) independently assessed each domain; disagreements were resolved by discussion.\\u0026nbsp;\\u003c/p\\u003e\\n\\u003cp\\u003e\\u003cem\\u003eData synthesis\\u003c/em\\u003e\\u003c/p\\u003e\\n\\u003cp\\u003eWe synthesized findings narratively, stratified by task category, disease focus, and LLM family. Within task categories, we grouped studies by specific infections (e.g., pneumonia/respiratory, bloodstream infection [BSI]/sepsis, catheter-associated urinary tract infection [CAUTI], invasive fungal infection [IFI/IMD], HIV/sexually transmitted infections [STIs], tuberculosis [TB]). When three or more commensurable studies examined the same task, disease, and metric, we reported descriptive statistics (median [IQR]) without formal meta-analytic pooling. \\u003c/p\\u003e\"},{\"header\":\"Results\",\"content\":\"\\u003cp\\u003e\\u003cem\\u003eStudy selection\\u003c/em\\u003e\\u003c/p\\u003e\\n\\u003cp\\u003eThe search retrieved 6,060 records from PubMed (1,329), CENTRAL (912), Scopus (3,379), and Web of Science (440). After removing 2,408 duplicates, 3,652 records were screened. Of these, 3,471 were excluded for irrelevance or pediatric focus. The remaining 181 full texts were reviewed, and 150 were excluded for not meeting inclusion criteria. Thirty-one studies, published between 2023 and 2025, were included in the final analysis (\\u003cstrong\\u003eFigure 1\\u003c/strong\\u003e).\\u003csup\\u003e3\\u0026ndash;33\\u003c/sup\\u003e\\u0026nbsp;\\u003c/p\\u003e\\n\\u003cp\\u003e\\u003cem\\u003eRisk of bias\\u003c/em\\u003e\\u003c/p\\u003e\\n\\u003cp\\u003eQUADAS-AI assessment revealed low overall risk in 7/31 studies (23%) and high risk in 24/31 (77%) (\\u003cstrong\\u003eTable Sup 1\\u003c/strong\\u003e). \\u003cstrong\\u003ePatient selection\\u003c/strong\\u003e showed high risk in 24 studies (77%), primarily attributable to reliance on synthetic vignettes or curated guideline-based questions rather than consecutive real patient encounters. \\u003cstrong\\u003eMeanwhile, the remaining domains exhibited predominantly low risk: index test (24/31 low risk, 77%), reference standard (27/31 low risk, 87%), and flow and timing (29/31 low risk, 94%).\\u003c/strong\\u003e\\u003c/p\\u003e\\n\\u003cp\\u003e\\u003cem\\u003eStudy characteristics\\u0026nbsp;\\u003c/em\\u003e\\u003c/p\\u003e\\n\\u003cp\\u003eThe 31 included studies evaluated LLM applications across diverse settings: inpatient wards (n=8), ICU (n=1), outpatient or primary care (n=1), mixed clinical settings (n=10), public health (n=2), laboratory (n=1), and comparative bench/question-based studies (n=8). Study designs comprised cross-sectional evaluations (n=19), prospective studies (n=5), retrospective analyses (n=4), and specialized accuracy/pilot studies (n=3).\\u003c/p\\u003e\\n\\u003cp\\u003eDisease foci included sexually transmitted diseases (n=4), urinary tract infections (n=4), tuberculosis (n=2), sepsis (n=2), HIV (n=2), hepatitis (n=2), mixed ID (n=3), and single studies on bacteremia, osteoarticular infections, vertebral osteomyelitis, blood culture diagnosis, surgical site infection, pneumonia, antimicrobial resistance, Infection Prevention and Control(IPC) guidelines, infective endocarditis prophylaxis, and general ID topics. Sample sizes ranged from 6 vignettes to 4,786 patients. \\u0026nbsp;\\u003c/p\\u003e\\n\\u003cp\\u003eLLM families included GPT-4/4o (n=20 studies), Claude 3/3.5 (n=7), Gemini (n=10), Mistral/Llama (n=3), custom chatbots (n=2), and specialized tools (n=6 studies using Consensus, OpenEvidence, Amboss, Perplexity, Microsoft Copilot). Retrieval-augmented generation or agentic architectures were explicitly implemented in 5 studies.\\u003c/p\\u003e\\n\\u003cp\\u003eData sources were vignettes/synthetic cases (n=21) and real EHR or clinical data (n=10). Comparators included infectious disease specialists (n=8), guidelines (n=8), other LLMs (n=13), or none (n=2) (\\u003cstrong\\u003eSupplementary\\u003c/strong\\u003e \\u003cstrong\\u003eTable 2\\u003c/strong\\u003e).\\u003c/p\\u003e\\n\\u003cp\\u003e\\u003cem\\u003eSafety and Clinical Applicability\\u003c/em\\u003e\\u003c/p\\u003e\\n\\u003cp\\u003eSafety concerns were documented in 28/31 studies (90.3%). Clinically incomplete responses represented the most prevalent issue (19/31, 61.3%), with usefulness ratings of 42% and mean completeness scores of 40%. Unsafe recommendations occurred in 10/31 studies (32.3%), manifesting as substandard responses posing serious health risk in 10.3% of evaluated cases. Harmful or inadequate recommendations appeared in 7/31 studies (22.6%): 16% potentially hazardous bacteremia management plans (Maillard et al.)\\u003csup\\u003e32\\u003c/sup\\u003e, 71% incorrect isolation precautions in sepsis cases (Lorenzoni et al.)\\u003csup\\u003e3\\u003c/sup\\u003e, and 9% inadequate source-control recommendations rates significantly exceeding infectious disease expert benchmarks (1\\u0026ndash;4%, p\\u0026lt;0.05). \\u003csup\\u003e32\\u003c/sup\\u003e\\u003c/p\\u003e\\n\\u003cp\\u003eHallucinations (fluent, confident response that presents incorrect information, akin to clinical confabulation), fabricated citations, contradictory statements, or artificial clinical details were documented in 10/31 studies (32.3%). Constraint-induced limitations (guardrail refusals, unanswered queries, token limits preventing processing of complete medical records) affected 9/31 studies (29%): 8% of interactions yielded no response and 65% experienced extraction errors from incomplete chart access. Context-dependent failures (errors mitigated by providing complete clinical data or external knowledge bases) in 3 studies, (3/31, 9.7%) demonstrated marked improvement with interventions: retrieval-augmented generation reduced hallucinations; providing complete clinical documentation improved central line-associated bloodstream infection detection specificity from 35% to 75% (Rodriguez-Nava et al.). These findings indicate requirements for expert oversight, structured validation protocols, and context-enrichment architectures before clinical deployment (Supplementary Table 2).\\u003c/p\\u003e\\n\\u003cp\\u003e\\u003cem\\u003eAntimicrobial Stewardship Performance\\u003c/em\\u003e\\u003c/p\\u003e\\n\\u003cp\\u003eTwelve studies assessed LLMs in antimicrobial stewardship. Across tasks, concordance with infectious-disease specialists was moderate, averaging about 50%. Two bacteremia vignette studies (n=100 each) reported identical 51% agreement (\\u0026kappa;=0.48). Agreement was higher for Gram-positive (70%, \\u0026kappa;=0.68) than Gram-negative infections (46%, \\u0026kappa;=0.43). In Maillard et al.\\u0026rsquo;s prospective study using real patient data (n=44, GPT-4), diagnostic accuracy reached 59% and empiric therapy appropriateness 64%, but 16% of recommendations were potentially harmful, including inactive agents and missed source-control interventions.\\u003csup\\u003e32\\u003c/sup\\u003e In a blood culture stewardship study (n = 84), LLMs produced 13% harmful or inadequate recommendations, significantly higher than experts (4%, p = 0.047). Most errors involved missing echocardiography for suspected endovascular infections.\\u003csup\\u003e24\\u003c/sup\\u003e\\u003c/p\\u003e\\n\\u003cp\\u003eA 14-model comparison showed wide variation in antibiotic prescribing accuracy. Antibiotic selection ranged from 30% to 100%, while dose and duration accuracy fell to 0\\u0026ndash;92%. Citation accuracy ranged 0\\u0026ndash;100%, and several models produced fabricated references. ChatGPT-o1 performed best overall (71.7% correct, 43/60).\\u003csup\\u003e30\\u003c/sup\\u003e In real-patient testing, Lorenzoni et al. (n = 7, GPT-4o) achieved perfect concordance for antibiotic selection but misjudged isolation precautions in 71% of cases.\\u003csup\\u003e3\\u003c/sup\\u003e Rodriguez-Nava and colleagues reported initially low central line-associated bloodstream infection (CLABSI) detection specificity (35%), which improved to 75% when complete chart information was provided, highlighting the critical role of RAG in improving accuracy.\\u003csup\\u003e11\\u003c/sup\\u003e In outpatient vignettes (n = 24, six models), Nguyen et al. found correct antibiotic selection between 59% and 100% and complete clinical advice between 25% and 96%, with proprietary models outperforming open-source ones.\\u003csup\\u003e29\\u003c/sup\\u003e\\u003c/p\\u003e\\n\\u003cp\\u003e\\u003cem\\u003eDiagnostic Accuracy and Guideline Concordance\\u003c/em\\u003e\\u003c/p\\u003e\\n\\u003cp\\u003eFour studies evaluated LLM diagnostic accuracy for infectious conditions using sensitivity and specificity metrics. Diagnostic sensitivity ranged from 80% to 100% (median 91%), with high performance for catheter-associated urinary tract infection detection (91%), tuberculous pleural effusion diagnosis (89%), and surgical site infection screening (100%). Specificity varied substantially (range 35-100%, median 92%). Wu and colleagues demonstrated that a custom LLM for tuberculous pleural effusion diagnosis achieved AUROC 0\\u0026middot;96 (sensitivity 76%, specificity 100%), matching or exceeding traditional machine learning models.\\u003csup\\u003e17\\u003c/sup\\u003e Satheakeerthy and colleagues showed zero-shot Llama-3-70B could screen surgical site infections with 100% sensitivity and 86% specificity, flagging infections earlier than infection control staff in 50% of cases.\\u003csup\\u003e12\\u003c/sup\\u003e\\u0026nbsp;\\u003c/p\\u003e\\n\\u003cp\\u003eTwelve studies evaluated LLMs in answering guideline-based questions, with accuracy ranging from 42% to 98% depending on topic and complexity. Borgonovo et al. found specialized RAG tools (Open Evidence, Microsoft Copilot) most accurate (94.4%), outperforming general-purpose models such as GPT-4o and Gemini 2.5 Pro (92.9%).\\u003csup\\u003e26\\u003c/sup\\u003e Lin et al. showed OpenAI O1 performed best for pneumonia guidelines, achieving 55% \\u0026ldquo;excellent\\u0026rdquo; responses and effective self-correction, compared with GPT-4o, which produced 40% \\u0026ldquo;poor\\u0026rdquo; responses.\\u003csup\\u003e20\\u003c/sup\\u003e Kufel et al. reported only 41.8% of GPT-3.5 outputs were rated \\u0026ldquo;useful,\\u0026rdquo; with low completeness (5.8/10) and safety (6.4/10).\\u003csup\\u003e28\\u003c/sup\\u003e Accuracy also varied by disease: tuberculosis questions scored 3.6\\u0026ndash;4.4/5, viral hepatitis 71\\u0026ndash;78%, and infective endocarditis prophylaxis 69\\u0026ndash;80%, with GPT-4o highest (80%). Performance was consistently better for informal, social-media-style questions (92%) than for formal guideline queries (69%, p \\u0026lt; 0.001) (\\u003cstrong\\u003eFigure 2, Supplementary Table 3.\\u003c/strong\\u003e).\\u003csup\\u003e9\\u003c/sup\\u003e\\u003c/p\\u003e\\n\\u003cp\\u003e\\u003cem\\u003eLimitations and Methodological Quality\\u003c/em\\u003e\\u003c/p\\u003e\\n\\u003cp\\u003eMethodological quality was constrained by design and reporting deficiencies. More than two-thirds of studies (21/31, 68%) evaluated LLMs using synthetic vignettes or curated cases, potentially overestimating performance by avoiding complexity, ambiguity, and incomplete documentation characteristic of actual practice. Among studies using real patient data (32%), most were retrospective with attendant selection bias. Sample sizes were frequently inadequate (35%), yielding unstable estimates. Validation approaches were weak across the evidence base. 12 studies (39%) had no expert comparison, instead benchmarking LLMs against other LLMs or guideline text alone. 14 studies (45%) judged correctness by guideline concordance without clinical context. The predominant cross-sectional design (22/31, 71%) precluded assessment of performance stability over time, while lack of blinding in 24/31 (77%) introduced expectation bias that might favor novel technology. Reproducibility and transparency were deficient. Most studies queried each question only once without assessing response variability, potentially concealing inconsistency and raising concerns about selective reporting of optimal responses. Process measures were absent in 21/31 (68%), leaving intervention fidelity and implementation strategies unclear. Outcome definitions were universally heterogeneous, employing non-standardized custom scales for usefulness, appropriateness, or quality that precluded cross-study comparison or meta-analysis (\\u003cstrong\\u003eTable 2\\u003c/strong\\u003e).\\u003c/p\\u003e\"},{\"header\":\"Discussion\",\"content\":\"\\u003cp\\u003eIn this systematic review of 31 studies, we found wide variability in how LLMs perform across infectious disease medical practice. Their accuracy and reliability remain insufficient for autonomous clinical use. Across antimicrobial stewardship, diagnosis, guideline interpretation, and surveillance tasks, performance was inconsistent. Concordance with infectious-disease specialists for empiric therapy averaged about Taken together, the evidence suggests that LLM diagnostic performance is strongest for narrow, well-specified classification tasks, ranging from common syndromes such as CAUTI,SSI to rarer entities such as tuberculous pleural effusions. However, accuracy and calibration consistently decline when models must integrate context across comorbidities, timelines, devices, cultures, and competing diagnoses, which is where clinical reasoning is most vulnerable to error. According to current evidence, RAG can improve specificity by providing access to full clinical documentation. Yet safety remains a central limitation: most studies reported incomplete or unsafe recommendations and, in many cases, fabricated or contradictory content. Even the most advanced models (at time of evaluation), such as GPT-4o and Claude 3.5, performed better than open-source systems but still failed to reach expert reliability. Their stronger results on conversational or social-media-style questions compared with formal guideline queries suggest training data skewed toward lay information rather than specialist clinical knowledge.\\u003c/p\\u003e\\n\\u003cp\\u003e\\u003cem\\u003eSafety and clinical applicability\\u003c/em\\u003e\\u003c/p\\u003e\\n\\u003cp\\u003eSafety has emerged as a significant barrier to the clinical deployment of LLMs in ID management. Rates of harmful or inadequate antimicrobial recommendations from these models ranged between 13% and 16%, sharply contrasting with the 1% to 4% error rates reported by infectious disease experts (Maillard et al.). Such harmful recommendations were described in Schwartz et al. document a paradigm case: when prompted to create a management plan for cryptococcal meningitis, GPT-3.5 recommended initiating antiretroviral therapy within 2 weeks-a recommendation\\u003csup\\u003e34\\u003c/sup\\u003e directly contradicted by Boulware et al.\\u0026apos;s randomized controlled trial proving this approach increases mortality.\\u003csup\\u003e35\\u003c/sup\\u003e Such inaccuracies pose serious risks by jeopardizing patient safety and exacerbating the global issue of AMR.\\u003csup\\u003e36\\u003c/sup\\u003e The underlying causes of these errors may mirror those observed in AMR challenges globally, primarily arising from various factors, such as wrong indication, selection, dosage, duration, lack of adherence to infection prevention and control (IPC) protocols.\\u003csup\\u003e37,38\\u003c/sup\\u003e LLMs showed this issues when they were lacking a defined guidelines, leading to generate incorrect responses. This challenge is further compounded by geographic variability in treatment protocols (Nguyen et al.); as established LLMs often lack access to localized medical knowledge.\\u003csup\\u003e39\\u003c/sup\\u003e This highlights the imperative for implementing RAG systems, which would enable LLMs to integrate context-specific information, thereby aligning with established protocols and mitigating the risk of adverse treatments. Another contributing factor to these inaccuracies is the inherent issue of hallucinations generated by LLMs.\\u003csup\\u003e40\\u003c/sup\\u003e While not errors per se (byproduct of LLMs training)\\u003csup\\u003e41\\u003c/sup\\u003e, these hallucinations manifesting as spurious guideline citations, contradictory assertions within a single interaction, or fabricated clinical details were documented in approximately one-third of the studies reviewed, undermining clinical trust and introducing medico-legal risks. This phenomenon is linked to LLMs\\u0026apos; limited access to real-time resources.\\u003csup\\u003e42\\u003c/sup\\u003e However, research indicates that providing web access significantly enhances their ability to generate accurate, high-quality scientific references.\\u003csup\\u003e43\\u003c/sup\\u003e Given these considerations, LLMs should not be deployed in their unmodified form due to their potential threats in the field (\\u003cstrong\\u003eFigure 3.\\u003c/strong\\u003e). Instead, they should be utilized as part of AI agents that leverage LLM capabilities while planning tasks, accessing external tools, and coordinating with other agents. In contrast to standard LLMs, these agents can perform multi-step processes, access real-time clinical information, and synthesize data from diverse sources.\\u003csup\\u003e44\\u003c/sup\\u003e This approach addresses the aforementioned safety concerns while also tackling additional issues such as verbosity, the need for expert-in-the-loop safety mechanisms, and iterative improvement. Due to the dearth of studies focusing specifically on ID, we cannot assertively conclude that such AI agents will resolve these challenges. However, this represents a critical area for future research as the next phase of LLM studies in infectious disease should aim to explore these innovations.\\u003c/p\\u003e\\n\\u003cp\\u003eCAUTI and CLABSI exemplify both the promise and the ceiling for LLM-enabled hospital epidemiology. These endpoints are operational quality metrics defined by NHSN surveillance rules,\\u003csup\\u003e45,46\\u003c/sup\\u003e not bedside diagnoses, so adjudication hinges on consistent application of standardized criteria to temporally ordered device, culture, and symptom data, often under substantial infection-prevention workload.\\u003csup\\u003e21\\u003c/sup\\u003e In CAUTI, GPT-4 achieved high performance on curated case abstractions and improved further with iterative, criteria-aligned prompting, highlighting how structured inputs and explicit rule framing can materially shift reliability. In contrast, CLABSI identification from real clinical notes was strongly context dependent: when the model was constrained to partial chart excerpts, sensitivity remained high but specificity was limited, and performance improved when key missing chart elements were supplied, supporting the use of RAG to pull the relevant EHR fields and NHSN rule elements before generation. Near term, LLMs are best deployed as definition-aware tools for education and structured abstraction that priorities sensitivity and trigger escalation, while final CAUTI and CLABSI attribution remains with trained infection-prevention reviewers. Because surveillance labels can propagate to antimicrobial decisions, CAUTI workflows should include explicit stewardship safeguards that confirm symptoms and exclude asymptomatic bacteriuria before outputs are acted on.\\u0026nbsp;\\u0026nbsp;\\u003c/p\\u003e\\n\\u003cp\\u003e\\u003cem\\u003eComparison with prior literature\\u003c/em\\u003e\\u003c/p\\u003e\\n\\u003cp\\u003eUp until now the topic of AI in ID is residing in a controversial place. A stark example is the contradiction between Siddig et al. declaring AI \\u0026quot;revolutionizing\\u0026quot; ID control\\u003csup\\u003e47\\u003c/sup\\u003e versus Schwartz\\u0026apos;s et al. \\u0026quot;Black Box Warning\\u0026quot; arguing \\u0026quot;existing LLMs are not safe for clinical consultation\\u0026rdquo;.\\u003csup\\u003e34\\u003c/sup\\u003e\\u0026nbsp;\\u003c/p\\u003e\\n\\u003cp\\u003eOur own group\\u0026rsquo;s previous systematic review of 15 studies identified promise of NLP and LLM in areas like pathogen detection and surveillance but noted limited real-world validation.\\u003csup\\u003e48\\u003c/sup\\u003e Our current findings extend beyond that review which included by the time of publications only two LLM studies. Be it as it may our current results show persistence of LLM limitation in literature, as evidenced by our observation that 58% of included studies relied on synthetic vignettes rather than real-world data.\\u003c/p\\u003e\\n\\u003cp\\u003eHoward et al. highlighted similar challenges in data completeness and interoperability for AI in tackling AMR, advocating for support of UN General Assembly targets like antimicrobial stewardship programs and surveillance, though with less focus on LLM-specific safety risks.\\u003csup\\u003e49\\u003c/sup\\u003e Despite this, our review found that LLMs may be useful in specific niches: diagnostics for urinary tract infections (UTIs), pneumonia, bloodstream infections (BSIs), and invasive fungal infections in defined populations (median AUROC 0.82, range 0.64\\u0026ndash;0.95); social media-based disease surveillance (accuracy 85\\u0026ndash;100%, with 92% for informal queries); antimicrobial stewardship (median appropriateness 71%, range 57\\u0026ndash;85%); and infection prevention/control with structured prompts (accuracy 98\\u0026ndash;100%). These successes share common features of well-defined diagnostic criteria, structured data sources, supplementary human verification, and consequences of errors that allow correction before patient harm.\\u003c/p\\u003e\\n\\u003cp\\u003e\\u003cem\\u003eLimitations\\u003c/em\\u003e\\u003c/p\\u003e\\n\\u003cp\\u003eOur systematic review has several limitations. First, substantial heterogeneity in populations, interventions, and outcome measures precluded meta-analysis. Second, the overall risk of bias was high across included studies. Third, the rapidly evolving nature of LLM technology means newer models might show different performance characteristics than those evaluated in our included studies. Finally, the predominance of retrospective studies in our review indicates a need for more prospective studies to validate the findings.\\u003c/p\\u003e\"},{\"header\":\"Conclusions\",\"content\":\"\\u003cp\\u003eThe reviewed iterations of LLMs are not well-suited for clinical application in ID. Most studies highlight safety concerns despite the models demonstrating high performance in structured tasks (UTI, BSI and fungal infections). Issues such as hallucinations, missing guideline information, and lack of web search capabilities contribute to misinformation from LLMs. To mitigate these challenges, LLMs should be employed as AI agents before being utilized in future studies.\\u003c/p\\u003e\"},{\"header\":\"Declarations\",\"content\":\"\\u003ch2\\u003eCompeting interests.\\u003c/h2\\u003e \\u003cp\\u003eThe authors declare that they have no competing interests.\\u003c/p\\u003e \\u003ch2\\u003eFunding.\\u003c/h2\\u003e \\u003cp\\u003eThis work was supported in part through the computational and data resources and staff expertise provided by Scientific Computing and Data at the Icahn School of Medicine at Mount Sinai and supported by the Clinical and Translational Science Awards (CTSA) grant UL1TR004419 from the National Center for Advancing Translational Sciences. Research reported in this publication was also supported by the Office of Research Infrastructure of the National Institutes of Health under award number S10OD026880 and S10OD030463. The content is solely the responsibility of the authors and does not necessarily represent the official views of the National Institutes of Health.\\u003c/p\\u003e\\u003ch2\\u003eAuthor Contributions:\\u003c/h2\\u003e \\u003cp\\u003eConceptualization, AG, MO, EK, BSG, GN; Methodology, AG, MO, EK; Formal Analysis, AG, MO; Data Curation, AG, MO; Writing-Original Draft Preparation, AG, MO; Writing- Review \\u0026amp; Editing, AG, MO, EK,MCP,RD,JJS, BSG,GN; Supervision: EK,GN\\u003c/p\\u003e\"},{\"header\":\"References\",\"content\":\"\\u003col\\u003e\\u003cli\\u003e\\u003cspan\\u003ePage MJ, McKenzie JE, Bossuyt PM et al (2021) The PRISMA 2020 statement: an updated guideline for reporting systematic reviews. BMJ 372:n71. \\u003cspan class=\\\"ExternalRef\\\"\\u003e\\u003cspan class=\\\"RefSource\\\"\\u003e10.1136/bmj.n71\\u003c/span\\u003e\\u003cspan address=\\\"10.1136/bmj.n71\\\" targettype=\\\"DOI\\\" class=\\\"RefTarget\\\"\\u003e\\u003c/span\\u003e\\u003c/span\\u003e\\u003c/span\\u003e\\u003c/li\\u003e \\u003cli\\u003e\\u003cspan\\u003eGuni A, Sounderajah V, Whiting P, Bossuyt P, Darzi A, Ashrafian H (2024) Revised Tool for the Quality Assessment of Diagnostic Accuracy Studies Using AI (QUADAS-AI): Protocol for a Qualitative Study. JMIR Res Protoc 13:e58202. \\u003cspan class=\\\"ExternalRef\\\"\\u003e\\u003cspan class=\\\"RefSource\\\"\\u003e10.2196/58202\\u003c/span\\u003e\\u003cspan address=\\\"10.2196/58202\\\" targettype=\\\"DOI\\\" class=\\\"RefTarget\\\"\\u003e\\u003c/span\\u003e\\u003c/span\\u003e\\u003c/span\\u003e\\u003c/li\\u003e \\u003cli\\u003e\\u003cspan\\u003eLorenzoni G, Garbin A, Brigiari G, Papappicco CAM, Manfrin V, Gregori D (2025) Large Language Models in Action: Supporting Clinical Evaluation in an Infectious Disease Unit. Healthcare 13(8):879. \\u003cspan class=\\\"ExternalRef\\\"\\u003e\\u003cspan class=\\\"RefSource\\\"\\u003e10.3390/healthcare13080879\\u003c/span\\u003e\\u003cspan address=\\\"10.3390/healthcare13080879\\\" targettype=\\\"DOI\\\" class=\\\"RefTarget\\\"\\u003e\\u003c/span\\u003e\\u003c/span\\u003e\\u003c/span\\u003e\\u003c/li\\u003e \\u003cli\\u003e\\u003cspan\\u003eEvaluation of artificial intelligence (AI) chatbots for providing sexual health information: a consensus study using real-world clinical queries | BMC Public Health | Full Text. Accessed October 8 (2025) \\u003cspan class=\\\"ExternalRef\\\"\\u003e\\u003cspan class=\\\"RefSource\\\"\\u003ehttps://bmcpublichealth.biomedcentral.com/articles/\\u003c/span\\u003e\\u003cspan address=\\\"https://bmcpublichealth.biomedcentral.com/articles/\\\" targettype=\\\"URL\\\" class=\\\"RefTarget\\\"\\u003e\\u003c/span\\u003e\\u003c/span\\u003e\\u003cspan class=\\\"ExternalRef\\\"\\u003e\\u003cspan class=\\\"RefSource\\\"\\u003e10.1186/s12889-025-22933-8\\u003c/span\\u003e\\u003cspan address=\\\"10.1186/s12889-025-22933-8\\\" targettype=\\\"DOI\\\" class=\\\"RefTarget\\\"\\u003e\\u003c/span\\u003e\\u003c/span\\u003e\\u003c/span\\u003e\\u003c/li\\u003e \\u003cli\\u003e\\u003cspan\\u003eDe Vito A, Geremia N, Marino A et al (2025) Assessing ChatGPT\\u0026rsquo;s theoretical knowledge and prescriptive accuracy in bacterial infections: a comparative study with infectious diseases residents and specialists. Infection 53(3):873\\u0026ndash;881. \\u003cspan class=\\\"ExternalRef\\\"\\u003e\\u003cspan class=\\\"RefSource\\\"\\u003e10.1007/s15010-024-02350-6\\u003c/span\\u003e\\u003cspan address=\\\"10.1007/s15010-024-02350-6\\\" targettype=\\\"DOI\\\" class=\\\"RefTarget\\\"\\u003e\\u003c/span\\u003e\\u003c/span\\u003e\\u003c/span\\u003e\\u003c/li\\u003e \\u003cli\\u003e\\u003cspan\\u003eAI Chatbots as Sources of STD Information: A Study on Reliability and Readability | Journal of Medical Systems. Accessed October 8 (2025) \\u003cspan class=\\\"ExternalRef\\\"\\u003e\\u003cspan class=\\\"RefSource\\\"\\u003ehttps://link.springer.com/article/\\u003c/span\\u003e\\u003cspan address=\\\"https://link.springer.com/article/\\\" targettype=\\\"URL\\\" class=\\\"RefTarget\\\"\\u003e\\u003c/span\\u003e\\u003c/span\\u003e\\u003cspan class=\\\"ExternalRef\\\"\\u003e\\u003cspan class=\\\"RefSource\\\"\\u003e10.1007/s10916-025-02178-z\\u003c/span\\u003e\\u003cspan address=\\\"10.1007/s10916-025-02178-z\\\" targettype=\\\"DOI\\\" class=\\\"RefTarget\\\"\\u003e\\u003c/span\\u003e\\u003c/span\\u003e\\u003c/span\\u003e\\u003c/li\\u003e \\u003cli\\u003e\\u003cspan\\u003eZhuo KY, Kim P, Kovacic J et al (2024) Can Artificial Intelligence Treat My Urinary Tract Infections?\\u0026mdash;Evaluation of Health Information Provided by OpenAI\\u003csup\\u003e\\u0026trade;\\u003c/sup\\u003e ChatGPT on Urinary Tract Infections. Soci\\u0026eacute;t\\u0026eacute; Int D\\u0026rsquo;Urologie J 5(2):104\\u0026ndash;107. \\u003cspan class=\\\"ExternalRef\\\"\\u003e\\u003cspan class=\\\"RefSource\\\"\\u003e10.3390/siuj5020018\\u003c/span\\u003e\\u003cspan address=\\\"10.3390/siuj5020018\\\" targettype=\\\"DOI\\\" class=\\\"RefTarget\\\"\\u003e\\u003c/span\\u003e\\u003c/span\\u003e\\u003c/span\\u003e\\u003c/li\\u003e \\u003cli\\u003e\\u003cspan\\u003eCakir H, Caglar U, Sekkeli S et al (2024) Evaluating ChatGPT ability to answer urinary tract Infection-Related questions. Infect Dis Now 54(4):104884. \\u003cspan class=\\\"ExternalRef\\\"\\u003e\\u003cspan class=\\\"RefSource\\\"\\u003e10.1016/j.idnow.2024.104884\\u003c/span\\u003e\\u003cspan address=\\\"10.1016/j.idnow.2024.104884\\\" targettype=\\\"DOI\\\" class=\\\"RefTarget\\\"\\u003e\\u003c/span\\u003e\\u003c/span\\u003e\\u003c/span\\u003e\\u003c/li\\u003e \\u003cli\\u003e\\u003cspan\\u003eRewthamrongsris P, Burapacheep J, Trachoo V, Porntaveetus T (2025) Accuracy of Large Language Models for Infective Endocarditis Prophylaxis in Dental Procedures. Int Dent J 75(1):206\\u0026ndash;212. \\u003cspan class=\\\"ExternalRef\\\"\\u003e\\u003cspan class=\\\"RefSource\\\"\\u003e10.1016/j.identj.2024.09.033\\u003c/span\\u003e\\u003cspan address=\\\"10.1016/j.identj.2024.09.033\\\" targettype=\\\"DOI\\\" class=\\\"RefTarget\\\"\\u003e\\u003c/span\\u003e\\u003c/span\\u003e\\u003c/span\\u003e\\u003c/li\\u003e \\u003cli\\u003e\\u003cspan\\u003eTun\\u0026ccedil;er G, G\\u0026uuml;\\u0026ccedil;l\\u0026uuml; KG (2024) How Reliable is ChatGPT as a Novel Consultant in Infectious Diseases and Clinical Microbiology? Infect Dis Clin Microbiol 6(1):55\\u0026ndash;59. \\u003cspan class=\\\"ExternalRef\\\"\\u003e\\u003cspan class=\\\"RefSource\\\"\\u003e10.36519/idcm.2024.286\\u003c/span\\u003e\\u003cspan address=\\\"10.36519/idcm.2024.286\\\" targettype=\\\"DOI\\\" class=\\\"RefTarget\\\"\\u003e\\u003c/span\\u003e\\u003c/span\\u003e\\u003c/span\\u003e\\u003c/li\\u003e \\u003cli\\u003e\\u003cspan\\u003eRodriguez-Nava G, Egoryan G, Goodman KE, Morgan DJ, Salinas JL (2024) Performance of a large language model for identifying central line-associated bloodstream infections (CLABSI) using real clinical notes. Infect Control Hosp Epidemiol 46(3):1\\u0026ndash;4. \\u003cspan class=\\\"ExternalRef\\\"\\u003e\\u003cspan class=\\\"RefSource\\\"\\u003e10.1017/ice.2024.164\\u003c/span\\u003e\\u003cspan address=\\\"10.1017/ice.2024.164\\\" targettype=\\\"DOI\\\" class=\\\"RefTarget\\\"\\u003e\\u003c/span\\u003e\\u003c/span\\u003e\\u003c/span\\u003e\\u003c/li\\u003e \\u003cli\\u003e\\u003cspan\\u003eSatheakeerthy S, Stretton B, Tsimiklis J et al Zero-shot large language model application for surgical site infection auditing. Infect Dis Health. Published online May 21, 2025:S2468-0451(25)00030\\u0026thinsp;\\u0026ndash;\\u0026thinsp;6. \\u003cspan class=\\\"ExternalRef\\\"\\u003e\\u003cspan class=\\\"RefSource\\\"\\u003e10.1016/j.idh.2025.05.001\\u003c/span\\u003e\\u003cspan address=\\\"10.1016/j.idh.2025.05.001\\\" targettype=\\\"DOI\\\" class=\\\"RefTarget\\\"\\u003e\\u003c/span\\u003e\\u003c/span\\u003e\\u003c/span\\u003e\\u003c/li\\u003e \\u003cli\\u003e\\u003cspan\\u003eCheah MH, Gan YN, Altice FL et al (2024) Testing the Feasibility and Acceptability of Using an Artificial Intelligence Chatbot to Promote HIV Testing and Pre-Exposure Prophylaxis in Malaysia: Mixed Methods Study. JMIR Hum Factors 11:e52055. \\u003cspan class=\\\"ExternalRef\\\"\\u003e\\u003cspan class=\\\"RefSource\\\"\\u003e10.2196/52055\\u003c/span\\u003e\\u003cspan address=\\\"10.2196/52055\\\" targettype=\\\"DOI\\\" class=\\\"RefTarget\\\"\\u003e\\u003c/span\\u003e\\u003c/span\\u003e\\u003c/span\\u003e\\u003c/li\\u003e \\u003cli\\u003e\\u003cspan\\u003eNi Z, Oh S, Saifi R, Azwa I, Altice FL (2025) Evaluating the Usability of an HIV Prevention Artificial Intelligence Chatbot in Malaysia: National Observational Study. JMIR Hum Factors 12:e70034. \\u003cspan class=\\\"ExternalRef\\\"\\u003e\\u003cspan class=\\\"RefSource\\\"\\u003e10.2196/70034\\u003c/span\\u003e\\u003cspan address=\\\"10.2196/70034\\\" targettype=\\\"DOI\\\" class=\\\"RefTarget\\\"\\u003e\\u003c/span\\u003e\\u003c/span\\u003e\\u003c/span\\u003e\\u003c/li\\u003e \\u003cli\\u003e\\u003cspan\\u003eLi Y, Huang CK, Hu Y, Zhou XD, He C, Zhong JW (2025) Exploring the performance of large language models on hepatitis B infection-related questions: A comparative study. World J Gastroenterol 31(3):101092. \\u003cspan class=\\\"ExternalRef\\\"\\u003e\\u003cspan class=\\\"RefSource\\\"\\u003e10.3748/wjg.v31.i3.101092\\u003c/span\\u003e\\u003cspan address=\\\"10.3748/wjg.v31.i3.101092\\\" targettype=\\\"DOI\\\" class=\\\"RefTarget\\\"\\u003e\\u003c/span\\u003e\\u003c/span\\u003e\\u003c/span\\u003e\\u003c/li\\u003e \\u003cli\\u003e\\u003cspan\\u003eComparison of the performances between ChatGPT and Gemini in answering questions on viral hepatitis | Scientific Reports. Accessed October 8 (2025) \\u003cspan class=\\\"ExternalRef\\\"\\u003e\\u003cspan class=\\\"RefSource\\\"\\u003ehttps://www.nature.com/articles/s41598-024-83575-1\\u003c/span\\u003e\\u003cspan address=\\\"https://www.nature.com/articles/s41598-024-83575-1\\\" targettype=\\\"URL\\\" class=\\\"RefTarget\\\"\\u003e\\u003c/span\\u003e\\u003c/span\\u003e\\u003c/span\\u003e\\u003c/li\\u003e \\u003cli\\u003e\\u003cspan\\u003eThe large language model diagnoses tuberculous pleural effusion in pleural effusion patients through clinical feature landscapes | Respiratory Research | Full Text. Accessed October 8 (2025) \\u003cspan class=\\\"ExternalRef\\\"\\u003e\\u003cspan class=\\\"RefSource\\\"\\u003ehttps://respiratory-research.biomedcentral.com/articles/\\u003c/span\\u003e\\u003cspan address=\\\"https://respiratory-research.biomedcentral.com/articles/\\\" targettype=\\\"URL\\\" class=\\\"RefTarget\\\"\\u003e\\u003c/span\\u003e\\u003c/span\\u003e\\u003cspan class=\\\"ExternalRef\\\"\\u003e\\u003cspan class=\\\"RefSource\\\"\\u003e10.1186/s12931-025-03130-y\\u003c/span\\u003e\\u003cspan address=\\\"10.1186/s12931-025-03130-y\\\" targettype=\\\"DOI\\\" class=\\\"RefTarget\\\"\\u003e\\u003c/span\\u003e\\u003c/span\\u003e\\u003c/span\\u003e\\u003c/li\\u003e \\u003cli\\u003e\\u003cspan\\u003eLarge (2025) language models\\u0026rsquo; capabilities in responding to tuberculosis medical questions: testing ChatGPT, Gemini, and Copilot | Scientific Reports. Accessed October 8. \\u003cspan class=\\\"ExternalRef\\\"\\u003e\\u003cspan class=\\\"RefSource\\\"\\u003ehttps://www.nature.com/articles/s41598-025-03074-9\\u003c/span\\u003e\\u003cspan address=\\\"https://www.nature.com/articles/s41598-025-03074-9\\\" targettype=\\\"URL\\\" class=\\\"RefTarget\\\"\\u003e\\u003c/span\\u003e\\u003c/span\\u003e\\u003c/span\\u003e\\u003c/li\\u003e \\u003cli\\u003e\\u003cspan\\u003eGiske CG, Bressan M, Fiechter F et al (2024) GPT-4-based AI agents-the new expert system for detection of antimicrobial resistance mechanisms? J Clin Microbiol 62(11):e0068924. \\u003cspan class=\\\"ExternalRef\\\"\\u003e\\u003cspan class=\\\"RefSource\\\"\\u003e10.1128/jcm.00689-24\\u003c/span\\u003e\\u003cspan address=\\\"10.1128/jcm.00689-24\\\" targettype=\\\"DOI\\\" class=\\\"RefTarget\\\"\\u003e\\u003c/span\\u003e\\u003c/span\\u003e\\u003c/span\\u003e\\u003c/li\\u003e \\u003cli\\u003e\\u003cspan\\u003eLin Z, Li Y, Wu M et al (2025) Performance analysis of large language models Chatgpt-4o, OpenAI O1, and OpenAI O3 mini in clinical treatment of pneumonia: a comparative study. Clin Exp Med 25(1):213. \\u003cspan class=\\\"ExternalRef\\\"\\u003e\\u003cspan class=\\\"RefSource\\\"\\u003e10.1007/s10238-025-01743-7\\u003c/span\\u003e\\u003cspan address=\\\"10.1007/s10238-025-01743-7\\\" targettype=\\\"DOI\\\" class=\\\"RefTarget\\\"\\u003e\\u003c/span\\u003e\\u003c/span\\u003e\\u003c/span\\u003e\\u003c/li\\u003e \\u003cli\\u003e\\u003cspan\\u003eWiemken TL, Carrico RM (2024) Assisting the infection preventionist: Use of artificial intelligence for health care-associated infection surveillance. Am J Infect Control 52(6):625\\u0026ndash;629. \\u003cspan class=\\\"ExternalRef\\\"\\u003e\\u003cspan class=\\\"RefSource\\\"\\u003e10.1016/j.ajic.2024.02.007\\u003c/span\\u003e\\u003cspan address=\\\"10.1016/j.ajic.2024.02.007\\\" targettype=\\\"DOI\\\" class=\\\"RefTarget\\\"\\u003e\\u003c/span\\u003e\\u003c/span\\u003e\\u003c/span\\u003e\\u003c/li\\u003e \\u003cli\\u003e\\u003cspan\\u003eAbosi OJ, Kobayashi T, Ross N et al (2024) A head-to-head comparison of the accuracy of commercially available large language models for infection prevention and control inquiries, 2024. Infect Control Hosp Epidemiol 46(3):1\\u0026ndash;3. \\u003cspan class=\\\"ExternalRef\\\"\\u003e\\u003cspan class=\\\"RefSource\\\"\\u003e10.1017/ice.2024.205\\u003c/span\\u003e\\u003cspan address=\\\"10.1017/ice.2024.205\\\" targettype=\\\"DOI\\\" class=\\\"RefTarget\\\"\\u003e\\u003c/span\\u003e\\u003c/span\\u003e\\u003c/span\\u003e\\u003c/li\\u003e \\u003cli\\u003e\\u003cspan\\u003eOh N, Cha WC, Seo JH et al (2024) ChatGPT Predicts In-Hospital All-Cause Mortality for Sepsis: In-Context Learning with the Korean Sepsis Alliance Database. Healthc Inf Res 30(3):266\\u0026ndash;276. \\u003cspan class=\\\"ExternalRef\\\"\\u003e\\u003cspan class=\\\"RefSource\\\"\\u003e10.4258/hir.2024.30.3.266\\u003c/span\\u003e\\u003cspan address=\\\"10.4258/hir.2024.30.3.266\\\" targettype=\\\"DOI\\\" class=\\\"RefTarget\\\"\\u003e\\u003c/span\\u003e\\u003c/span\\u003e\\u003c/span\\u003e\\u003c/li\\u003e \\u003cli\\u003e\\u003cspan\\u003eTassone DM, Hitchcock MM, Rossier CJ et al (2025) Evaluating chain-of-thought prompting in a GPT chatbot for BCID2 interpretation and stewardship: how does AI compare to human experts? Antimicrob Steward Healthc Epidemiol ASHE 5(1):e154. \\u003cspan class=\\\"ExternalRef\\\"\\u003e\\u003cspan class=\\\"RefSource\\\"\\u003e10.1017/ash.2025.10059\\u003c/span\\u003e\\u003cspan address=\\\"10.1017/ash.2025.10059\\\" targettype=\\\"DOI\\\" class=\\\"RefTarget\\\"\\u003e\\u003c/span\\u003e\\u003c/span\\u003e\\u003c/span\\u003e\\u003c/li\\u003e \\u003cli\\u003e\\u003cspan\\u003eMilicevic F, Ghandour M, Khasawneh MY et al (2025) Assessing LLMs on IDSA Practice Guidelines for the Diagnosis and Treatment of Native Vertebral Osteomyelitis: A Comparison Study. J Clin Med 14(14):4996. \\u003cspan class=\\\"ExternalRef\\\"\\u003e\\u003cspan class=\\\"RefSource\\\"\\u003e10.3390/jcm14144996\\u003c/span\\u003e\\u003cspan address=\\\"10.3390/jcm14144996\\\" targettype=\\\"DOI\\\" class=\\\"RefTarget\\\"\\u003e\\u003c/span\\u003e\\u003c/span\\u003e\\u003c/span\\u003e\\u003c/li\\u003e \\u003cli\\u003e\\u003cspan\\u003eBorgonovo F, Matsuo T, Petri F et al (2025) Battle of the Bots: Solving Clinical Cases in Osteoarticular Infections With Large Language Models. Mayo Clin Proc Digit Health 3(3):100230. \\u003cspan class=\\\"ExternalRef\\\"\\u003e\\u003cspan class=\\\"RefSource\\\"\\u003e10.1016/j.mcpdig.2025.100230\\u003c/span\\u003e\\u003cspan address=\\\"10.1016/j.mcpdig.2025.100230\\\" targettype=\\\"DOI\\\" class=\\\"RefTarget\\\"\\u003e\\u003c/span\\u003e\\u003c/span\\u003e\\u003c/span\\u003e\\u003c/li\\u003e \\u003cli\\u003e\\u003cspan\\u003eEvaluating the Performance of State-of-the-Art Artificial Intelligence Chatbots Based on the WHO Global Guidelines for the Prevention of Surgical Site Infection: Cross-Sectional Study - PubMed. Accessed October 8 (2025) \\u003cspan class=\\\"ExternalRef\\\"\\u003e\\u003cspan class=\\\"RefSource\\\"\\u003ehttps://pubmed.ncbi.nlm.nih.gov/40744114/\\u003c/span\\u003e\\u003cspan address=\\\"https://pubmed.ncbi.nlm.nih.gov/40744114/\\\" targettype=\\\"URL\\\" class=\\\"RefTarget\\\"\\u003e\\u003c/span\\u003e\\u003c/span\\u003e\\u003c/span\\u003e\\u003c/li\\u003e \\u003cli\\u003e\\u003cspan\\u003eKufel WD, Hanrahan KD, Seabury RW et al (2024) Let\\u0026rsquo;s Have a Chat: How Well Does an Artificial Intelligence Chatbot Answer Clinical Infectious Diseases Pharmacotherapy Questions? Open Forum Infect Dis 11(11):ofae641. \\u003cspan class=\\\"ExternalRef\\\"\\u003e\\u003cspan class=\\\"RefSource\\\"\\u003e10.1093/ofid/ofae641\\u003c/span\\u003e\\u003cspan address=\\\"10.1093/ofid/ofae641\\\" targettype=\\\"DOI\\\" class=\\\"RefTarget\\\"\\u003e\\u003c/span\\u003e\\u003c/span\\u003e\\u003c/span\\u003e\\u003c/li\\u003e \\u003cli\\u003e\\u003cspan\\u003eNgoc Nguyen O, Amin D, Bennett J et al (2025) GP or ChatGPT? Ability of large language models (LLMs) to support general practitioners when prescribing antibiotics. J Antimicrob Chemother 80(5):1324\\u0026ndash;1330. \\u003cspan class=\\\"ExternalRef\\\"\\u003e\\u003cspan class=\\\"RefSource\\\"\\u003e10.1093/jac/dkaf077\\u003c/span\\u003e\\u003cspan address=\\\"10.1093/jac/dkaf077\\\" targettype=\\\"DOI\\\" class=\\\"RefTarget\\\"\\u003e\\u003c/span\\u003e\\u003c/span\\u003e\\u003c/span\\u003e\\u003c/li\\u003e \\u003cli\\u003e\\u003cspan\\u003eDe Vito A, Geremia N, Bavaro DF et al (2025) Comparing large language models for antibiotic prescribing in different clinical scenarios: which performs better? Clin Microbiol Infect Off Publ Eur Soc Clin Microbiol Infect Dis 31(8):1336\\u0026ndash;1342. \\u003cspan class=\\\"ExternalRef\\\"\\u003e\\u003cspan class=\\\"RefSource\\\"\\u003e10.1016/j.cmi.2025.03.002\\u003c/span\\u003e\\u003cspan address=\\\"10.1016/j.cmi.2025.03.002\\\" targettype=\\\"DOI\\\" class=\\\"RefTarget\\\"\\u003e\\u003c/span\\u003e\\u003c/span\\u003e\\u003c/span\\u003e\\u003c/li\\u003e \\u003cli\\u003e\\u003cspan\\u003eMontiel-Romero S, Rajme-L\\u0026oacute;pez S, Rom\\u0026aacute;n-Montes CM et al (2025) Recommended antibiotic treatment agreement between infectious diseases specialists and ChatGPT\\u0026reg;. BMC Infect Dis 25(1):38. \\u003cspan class=\\\"ExternalRef\\\"\\u003e\\u003cspan class=\\\"RefSource\\\"\\u003e10.1186/s12879-024-10426-9\\u003c/span\\u003e\\u003cspan address=\\\"10.1186/s12879-024-10426-9\\\" targettype=\\\"DOI\\\" class=\\\"RefTarget\\\"\\u003e\\u003c/span\\u003e\\u003c/span\\u003e\\u003c/span\\u003e\\u003c/li\\u003e \\u003cli\\u003e\\u003cspan\\u003eMaillard A, Micheli G, Lefevre L et al (2024) Can Chatbot Artificial Intelligence Replace Infectious Diseases Physicians in the Management of Bloodstream Infections? A Prospective Cohort Study. Clin Infect Dis Off Publ Infect Dis Soc Am 78(4):825\\u0026ndash;832. \\u003cspan class=\\\"ExternalRef\\\"\\u003e\\u003cspan class=\\\"RefSource\\\"\\u003e10.1093/cid/ciad632\\u003c/span\\u003e\\u003cspan address=\\\"10.1093/cid/ciad632\\\" targettype=\\\"DOI\\\" class=\\\"RefTarget\\\"\\u003e\\u003c/span\\u003e\\u003c/span\\u003e\\u003c/span\\u003e\\u003c/li\\u003e \\u003cli\\u003e\\u003cspan\\u003ePerret J, Schmid A (2024) Application of OpenAI GPT-4 for the retrospective detection of catheter-associated urinary tract infections in a fictitious and curated patient data set. Infect Control Hosp Epidemiol 45(1):96\\u0026ndash;99. \\u003cspan class=\\\"ExternalRef\\\"\\u003e\\u003cspan class=\\\"RefSource\\\"\\u003e10.1017/ice.2023.189\\u003c/span\\u003e\\u003cspan address=\\\"10.1017/ice.2023.189\\\" targettype=\\\"DOI\\\" class=\\\"RefTarget\\\"\\u003e\\u003c/span\\u003e\\u003c/span\\u003e\\u003c/span\\u003e\\u003c/li\\u003e \\u003cli\\u003e\\u003cspan\\u003eSchwartz IS, Link KE, Daneshjou R, Cort\\u0026eacute;s-Penfield N (2023) Black Box Warning: Large Language Models and the Future of Infectious Diseases Consultation. Clin Infect Dis Off Publ Infect Dis Soc Am 78(4):860\\u0026ndash;866. \\u003cspan class=\\\"ExternalRef\\\"\\u003e\\u003cspan class=\\\"RefSource\\\"\\u003e10.1093/cid/ciad633\\u003c/span\\u003e\\u003cspan address=\\\"10.1093/cid/ciad633\\\" targettype=\\\"DOI\\\" class=\\\"RefTarget\\\"\\u003e\\u003c/span\\u003e\\u003c/span\\u003e\\u003c/span\\u003e\\u003c/li\\u003e \\u003cli\\u003e\\u003cspan\\u003eTiming of Antiretroviral Therapy after Diagnosis of Cryptococcal Meningitis | New England Journal of Medicine. Accessed October 7 (2025) \\u003cspan class=\\\"ExternalRef\\\"\\u003e\\u003cspan class=\\\"RefSource\\\"\\u003ehttps://www.nejm.org/doi/full/\\u003c/span\\u003e\\u003cspan address=\\\"https://www.nejm.org/doi/full/\\\" targettype=\\\"URL\\\" class=\\\"RefTarget\\\"\\u003e\\u003c/span\\u003e\\u003c/span\\u003e\\u003cspan class=\\\"ExternalRef\\\"\\u003e\\u003cspan class=\\\"RefSource\\\"\\u003e10.1056/NEJMoa1312884\\u003c/span\\u003e\\u003cspan address=\\\"10.1056/NEJMoa1312884\\\" targettype=\\\"DOI\\\" class=\\\"RefTarget\\\"\\u003e\\u003c/span\\u003e\\u003c/span\\u003e\\u003c/span\\u003e\\u003c/li\\u003e \\u003cli\\u003e\\u003cspan\\u003eGBD 2021 Antimicrobial Resistance Collaborators (2024) Global burden of bacterial antimicrobial resistance 1990\\u0026ndash;2021: a systematic analysis with forecasts to 2050. Lancet Lond Engl 404(10459):1199\\u0026ndash;1226. \\u003cspan class=\\\"ExternalRef\\\"\\u003e\\u003cspan class=\\\"RefSource\\\"\\u003e10.1016/S0140-6736(24)01867-1\\u003c/span\\u003e\\u003cspan address=\\\"10.1016/S0140-6736(24)01867-1\\\" targettype=\\\"DOI\\\" class=\\\"RefTarget\\\"\\u003e\\u003c/span\\u003e\\u003c/span\\u003e\\u003c/span\\u003e\\u003c/li\\u003e \\u003cli\\u003e\\u003cspan\\u003eOtaigbe II, Elikwu CJ (2023) Drivers of inappropriate antibiotic use in low- and middle-income countries. JAC-Antimicrob Resist 5(3):dlad062. \\u003cspan class=\\\"ExternalRef\\\"\\u003e\\u003cspan class=\\\"RefSource\\\"\\u003e10.1093/jacamr/dlad062\\u003c/span\\u003e\\u003cspan address=\\\"10.1093/jacamr/dlad062\\\" targettype=\\\"DOI\\\" class=\\\"RefTarget\\\"\\u003e\\u003c/span\\u003e\\u003c/span\\u003e\\u003c/span\\u003e\\u003c/li\\u003e \\u003cli\\u003e\\u003cspan\\u003eMulchandani R, Tiseo K, Nandi A et al (2025) Global trends in inappropriate use of antibiotics, 2000\\u0026ndash;2021: scoping review and prevalence estimates. BMJ Public Health 3(1):e002411. \\u003cspan class=\\\"ExternalRef\\\"\\u003e\\u003cspan class=\\\"RefSource\\\"\\u003e10.1136/bmjph-2024-002411\\u003c/span\\u003e\\u003cspan address=\\\"10.1136/bmjph-2024-002411\\\" targettype=\\\"DOI\\\" class=\\\"RefTarget\\\"\\u003e\\u003c/span\\u003e\\u003c/span\\u003e\\u003c/span\\u003e\\u003c/li\\u003e \\u003cli\\u003e\\u003cspan\\u003eBusch F, Hoffmann L, Rueger C et al (2025) Current applications and challenges in large language models for patient care: a systematic review. Commun Med 5(1):1\\u0026ndash;13. \\u003cspan class=\\\"ExternalRef\\\"\\u003e\\u003cspan class=\\\"RefSource\\\"\\u003e10.1038/s43856-024-00717-2\\u003c/span\\u003e\\u003cspan address=\\\"10.1038/s43856-024-00717-2\\\" targettype=\\\"DOI\\\" class=\\\"RefTarget\\\"\\u003e\\u003c/span\\u003e\\u003c/span\\u003e\\u003c/span\\u003e\\u003c/li\\u003e \\u003cli\\u003e\\u003cspan\\u003eOmar M, Sorin V, Collins JD et al (2025) Multi-model assurance analysis showing large language models are highly vulnerable to adversarial hallucination attacks during clinical decision support. Commun Med 5(1):330. \\u003cspan class=\\\"ExternalRef\\\"\\u003e\\u003cspan class=\\\"RefSource\\\"\\u003e10.1038/s43856-025-01021-3\\u003c/span\\u003e\\u003cspan address=\\\"10.1038/s43856-025-01021-3\\\" targettype=\\\"DOI\\\" class=\\\"RefTarget\\\"\\u003e\\u003c/span\\u003e\\u003c/span\\u003e\\u003c/span\\u003e\\u003c/li\\u003e \\u003cli\\u003e\\u003cspan\\u003eKalai AT, Nachum O, Vempala SS, Zhang E (2025) Why Language Models Hallucinate. \\u003cem\\u003earXiv\\u003c/em\\u003e. Preprint posted online September 4. \\u003cspan class=\\\"ExternalRef\\\"\\u003e\\u003cspan class=\\\"RefSource\\\"\\u003e10.48550/arXiv.2509.04664\\u003c/span\\u003e\\u003cspan address=\\\"10.48550/arXiv.2509.04664\\\" targettype=\\\"DOI\\\" class=\\\"RefTarget\\\"\\u003e\\u003c/span\\u003e\\u003c/span\\u003e\\u003c/span\\u003e\\u003c/li\\u003e \\u003cli\\u003e\\u003cspan\\u003eGenerating credible referenced medical research: A comparative study of openAI\\u0026rsquo;s GPT-4 and Google\\u0026rsquo;s gemini - ScienceDirect. Accessed December 16 (2024) \\u003cspan class=\\\"ExternalRef\\\"\\u003e\\u003cspan class=\\\"RefSource\\\"\\u003ehttps://www.sciencedirect.com/science/article/pii/S0010482524016305\\u003c/span\\u003e\\u003cspan address=\\\"https://www.sciencedirect.com/science/article/pii/S0010482524016305\\\" targettype=\\\"URL\\\" class=\\\"RefTarget\\\"\\u003e\\u003c/span\\u003e\\u003c/span\\u003e\\u003c/span\\u003e\\u003c/li\\u003e \\u003cli\\u003e\\u003cspan\\u003eGorenshtein A, Shihada K, Sorka M, Aran D, Shelly S (2025) LITERAS: Biomedical literature review and citation retrieval agents. Comput Biol Med 192:110363. \\u003cspan class=\\\"ExternalRef\\\"\\u003e\\u003cspan class=\\\"RefSource\\\"\\u003e10.1016/j.compbiomed.2025.110363\\u003c/span\\u003e\\u003cspan address=\\\"10.1016/j.compbiomed.2025.110363\\\" targettype=\\\"DOI\\\" class=\\\"RefTarget\\\"\\u003e\\u003c/span\\u003e\\u003c/span\\u003e\\u003c/span\\u003e\\u003c/li\\u003e \\u003cli\\u003e\\u003cspan\\u003eGorenshtein A, Omar M, Glicksberg BS, Nadkarni GN, Klang E AI Agents in Clinical Medicine: A Systematic Review. \\u003cem\\u003emedRxiv\\u003c/em\\u003e. Preprint posted online August 26, 2025:2025.08.22.25334232. \\u003cspan class=\\\"ExternalRef\\\"\\u003e\\u003cspan class=\\\"RefSource\\\"\\u003e10.1101/2025.08.22.25334232\\u003c/span\\u003e\\u003cspan address=\\\"10.1101/2025.08.22.25334232\\\" targettype=\\\"DOI\\\" class=\\\"RefTarget\\\"\\u003e\\u003c/span\\u003e\\u003c/span\\u003e\\u003c/span\\u003e\\u003c/li\\u003e \\u003cli\\u003e\\u003cspan\\u003eBloodstream Infections Published online 2025\\u003c/span\\u003e\\u003c/li\\u003e \\u003cli\\u003e\\u003cspan\\u003eUrinary Tract Infection Published online 2025\\u003c/span\\u003e\\u003c/li\\u003e \\u003cli\\u003e\\u003cspan\\u003eSiddig EE, Eltigani HF, Ahmed A (2023) The Rise of AI: How Artificial Intelligence is Revolutionizing Infectious Disease Control. Ann Biomed Eng 51(12):2636\\u0026ndash;2637. \\u003cspan class=\\\"ExternalRef\\\"\\u003e\\u003cspan class=\\\"RefSource\\\"\\u003e10.1007/s10439-023-03280-4\\u003c/span\\u003e\\u003cspan address=\\\"10.1007/s10439-023-03280-4\\\" targettype=\\\"DOI\\\" class=\\\"RefTarget\\\"\\u003e\\u003c/span\\u003e\\u003c/span\\u003e\\u003c/span\\u003e\\u003c/li\\u003e \\u003cli\\u003e\\u003cspan\\u003eOmar M, Brin D, Glicksberg B, Klang E (2024) Utilizing natural language processing and large language models in the diagnosis and prediction of infectious diseases: A systematic review. Am J Infect Control 52(9):992\\u0026ndash;1001. \\u003cspan class=\\\"ExternalRef\\\"\\u003e\\u003cspan class=\\\"RefSource\\\"\\u003e10.1016/j.ajic.2024.03.016\\u003c/span\\u003e\\u003cspan address=\\\"10.1016/j.ajic.2024.03.016\\\" targettype=\\\"DOI\\\" class=\\\"RefTarget\\\"\\u003e\\u003c/span\\u003e\\u003c/span\\u003e\\u003c/span\\u003e\\u003c/li\\u003e \\u003cli\\u003e\\u003cspan\\u003eHoward A, Reza N, Green PL et al (2025) Artificial intelligence and infectious diseases: tackling antimicrobial resistance, from personalised care to antibiotic discovery. Lancet Infect Dis 0(0). \\u003cspan class=\\\"ExternalRef\\\"\\u003e\\u003cspan class=\\\"RefSource\\\"\\u003e10.1016/S1473-3099(25)00313-5\\u003c/span\\u003e\\u003cspan address=\\\"10.1016/S1473-3099(25)00313-5\\\" targettype=\\\"DOI\\\" class=\\\"RefTarget\\\"\\u003e\\u003c/span\\u003e\\u003c/span\\u003e\\u003c/span\\u003e\\u003c/li\\u003e\\u003c/ol\\u003e\"},{\"header\":\"Tables\",\"content\":\"\\u003cp\\u003eTable 1. Safety issue categories, prevalence across studies, quantitative signals, and mitigations\\u003c/p\\u003e\\n\\u003ctable border=\\\"1\\\" cellspacing=\\\"0\\\" cellpadding=\\\"0\\\" width=\\\"794\\\"\\u003e\\n \\u003ctbody\\u003e\\n \\u003ctr\\u003e\\n \\u003ctd valign=\\\"top\\\" style=\\\"width: 16.1695%;\\\"\\u003e\\n \\u003cp\\u003e\\u003cstrong\\u003eSafety issue category\\u003c/strong\\u003e\\u003c/p\\u003e\\n \\u003c/td\\u003e\\n \\u003ctd valign=\\\"top\\\" style=\\\"width: 6.894%;\\\"\\u003e\\n \\u003cp\\u003e\\u003cstrong\\u003eStudies (n/31; %)\\u003c/strong\\u003e\\u003c/p\\u003e\\n \\u003c/td\\u003e\\n \\u003ctd valign=\\\"top\\\" style=\\\"width: 12.9105%;\\\"\\u003e\\n \\u003cp\\u003e\\u003cstrong\\u003eRepresentative quantitative rate(s)\\u003c/strong\\u003e\\u003c/p\\u003e\\n \\u003c/td\\u003e\\n \\u003ctd valign=\\\"top\\\" style=\\\"width: 7.7714%;\\\"\\u003e\\n \\u003cp\\u003e\\u003cstrong\\u003eSeverity tier\\u003c/strong\\u003e\\u003c/p\\u003e\\n \\u003c/td\\u003e\\n \\u003ctd valign=\\\"top\\\" style=\\\"width: 10.0276%;\\\"\\u003e\\n \\u003cp\\u003e\\u003cstrong\\u003eContext-dependent?\\u003c/strong\\u003e\\u003c/p\\u003e\\n \\u003c/td\\u003e\\n \\u003ctd valign=\\\"top\\\" style=\\\"width: 11.4064%;\\\"\\u003e\\n \\u003cp\\u003e\\u003cstrong\\u003eClinical risk domain(s)\\u003c/strong\\u003e\\u003c/p\\u003e\\n \\u003c/td\\u003e\\n \\u003ctd valign=\\\"top\\\" style=\\\"width: 10.0276%;\\\"\\u003e\\n \\u003cp\\u003e\\u003cstrong\\u003eCommon mitigations\\u003c/strong\\u003e\\u003c/p\\u003e\\n \\u003c/td\\u003e\\n \\u003ctd valign=\\\"top\\\" style=\\\"width: 136px;\\\"\\u003e\\n \\u003cp\\u003e\\u003cstrong\\u003eExample studies\\u003c/strong\\u003e\\u003c/p\\u003e\\n \\u003c/td\\u003e\\n \\u003c/tr\\u003e\\n \\u003ctr\\u003e\\n \\u003ctd valign=\\\"top\\\" style=\\\"width: 16.1695%;\\\"\\u003e\\n \\u003cp\\u003eUnsafe recommendations\\u003c/p\\u003e\\n \\u003c/td\\u003e\\n \\u003ctd valign=\\\"top\\\" style=\\\"width: 6.894%;\\\"\\u003e\\n \\u003cp\\u003e10/31 (32.3%)\\u003c/p\\u003e\\n \\u003c/td\\u003e\\n \\u003ctd valign=\\\"top\\\" style=\\\"width: 12.9105%;\\\"\\u003e\\n \\u003cp\\u003e7/68 (10.3%) substandard responses posing serious health risk\\u003c/p\\u003e\\n \\u003c/td\\u003e\\n \\u003ctd valign=\\\"top\\\" style=\\\"width: 7.7714%;\\\"\\u003e\\n \\u003cp\\u003eHigh\\u003c/p\\u003e\\n \\u003c/td\\u003e\\n \\u003ctd valign=\\\"top\\\" style=\\\"width: 10.0276%;\\\"\\u003e\\n \\u003cp\\u003eYes\\u003c/p\\u003e\\n \\u003c/td\\u003e\\n \\u003ctd valign=\\\"top\\\" style=\\\"width: 11.4064%;\\\"\\u003e\\n \\u003cp\\u003eGuideline adherence, dosing, antimicrobial choice, AMR interpretation\\u003c/p\\u003e\\n \\u003c/td\\u003e\\n \\u003ctd valign=\\\"top\\\" style=\\\"width: 10.0276%;\\\"\\u003e\\n \\u003cp\\u003eExpert review, external validation\\u003c/p\\u003e\\n \\u003c/td\\u003e\\n \\u003ctd valign=\\\"top\\\" style=\\\"width: 136px;\\\"\\u003e\\n \\u003cp\\u003eAbosi 2024, Cakir 2023\\u003c/p\\u003e\\n \\u003c/td\\u003e\\n \\u003c/tr\\u003e\\n \\u003ctr\\u003e\\n \\u003ctd valign=\\\"top\\\" style=\\\"width: 16.1695%;\\\"\\u003e\\n \\u003cp\\u003eHarmful/inadequate recommendations\\u003c/p\\u003e\\n \\u003c/td\\u003e\\n \\u003ctd valign=\\\"top\\\" style=\\\"width: 6.894%;\\\"\\u003e\\n \\u003cp\\u003e7/31 (22.6%)\\u003c/p\\u003e\\n \\u003c/td\\u003e\\n \\u003ctd valign=\\\"top\\\" style=\\\"width: 12.9105%;\\\"\\u003e\\n \\u003cp\\u003e7/44 (15.9%) harmful management plans; 5/7 (71.4%) incorrect isolation precautions; 4/44 (9.1%) inadequate source control\\u003c/p\\u003e\\n \\u003c/td\\u003e\\n \\u003ctd valign=\\\"top\\\" style=\\\"width: 7.7714%;\\\"\\u003e\\n \\u003cp\\u003eHigh\\u003c/p\\u003e\\n \\u003c/td\\u003e\\n \\u003ctd valign=\\\"top\\\" style=\\\"width: 10.0276%;\\\"\\u003e\\n \\u003cp\\u003eYes\\u003c/p\\u003e\\n \\u003c/td\\u003e\\n \\u003ctd valign=\\\"top\\\" style=\\\"width: 11.4064%;\\\"\\u003e\\n \\u003cp\\u003eIsolation/IPC, antimicrobial choice, dosing, source control\\u003c/p\\u003e\\n \\u003c/td\\u003e\\n \\u003ctd valign=\\\"top\\\" style=\\\"width: 10.0276%;\\\"\\u003e\\n \\u003cp\\u003ePrompt engineering, chain-of-thought prompting, expert review\\u003c/p\\u003e\\n \\u003c/td\\u003e\\n \\u003ctd valign=\\\"top\\\" style=\\\"width: 136px;\\\"\\u003e\\n \\u003cp\\u003eLorenzoni 2025, Maillard 2024\\u003c/p\\u003e\\n \\u003c/td\\u003e\\n \\u003c/tr\\u003e\\n \\u003ctr\\u003e\\n \\u003ctd valign=\\\"top\\\" style=\\\"width: 16.1695%;\\\"\\u003e\\n \\u003cp\\u003eHallucinations / fabricated citations\\u003c/p\\u003e\\n \\u003c/td\\u003e\\n \\u003ctd valign=\\\"top\\\" style=\\\"width: 6.894%;\\\"\\u003e\\n \\u003cp\\u003e10/31 (32.3%)\\u003c/p\\u003e\\n \\u003c/td\\u003e\\n \\u003ctd valign=\\\"top\\\" style=\\\"width: 12.9105%;\\\"\\u003e\\n \\u003cp\\u003enot reported\\u003c/p\\u003e\\n \\u003c/td\\u003e\\n \\u003ctd valign=\\\"top\\\" style=\\\"width: 7.7714%;\\\"\\u003e\\n \\u003cp\\u003eModerate\\u003c/p\\u003e\\n \\u003c/td\\u003e\\n \\u003ctd valign=\\\"top\\\" style=\\\"width: 10.0276%;\\\"\\u003e\\n \\u003cp\\u003eYes\\u003c/p\\u003e\\n \\u003c/td\\u003e\\n \\u003ctd valign=\\\"top\\\" style=\\\"width: 11.4064%;\\\"\\u003e\\n \\u003cp\\u003eAntimicrobial choice, dosing, citations\\u003c/p\\u003e\\n \\u003c/td\\u003e\\n \\u003ctd valign=\\\"top\\\" style=\\\"width: 10.0276%;\\\"\\u003e\\n \\u003cp\\u003eRAG to guidelines, external knowledge bases\\u003c/p\\u003e\\n \\u003c/td\\u003e\\n \\u003ctd valign=\\\"top\\\" style=\\\"width: 136px;\\\"\\u003e\\n \\u003cp\\u003eBorgonovo 2025, Cakir 2023\\u003c/p\\u003e\\n \\u003c/td\\u003e\\n \\u003c/tr\\u003e\\n \\u003ctr\\u003e\\n \\u003ctd colspan=\\\"4\\\" valign=\\\"top\\\" style=\\\"width: 43.7453%;\\\"\\u003e\\n \\u003cp\\u003e\\u003cstrong\\u003eInadequate or incomplete responses\\u003c/strong\\u003e\\u003c/p\\u003e\\n \\u003c/td\\u003e\\n \\u003ctd valign=\\\"top\\\" style=\\\"width: 10.0276%;\\\"\\u003e\\u003cbr\\u003e\\u003c/td\\u003e\\n \\u003ctd valign=\\\"top\\\" style=\\\"width: 11.4064%;\\\"\\u003e\\u003cbr\\u003e\\u003c/td\\u003e\\n \\u003ctd valign=\\\"top\\\" style=\\\"width: 10.0276%;\\\"\\u003e\\u003cbr\\u003e\\u003c/td\\u003e\\n \\u003ctd valign=\\\"top\\\" style=\\\"width: 10.6543%;\\\"\\u003e\\n \\u003cp\\u003e\\u0026nbsp;\\u003c/p\\u003e\\n \\u003c/td\\u003e\\n \\u003c/tr\\u003e\\n \\u003ctr\\u003e\\n \\u003ctd valign=\\\"top\\\" style=\\\"width: 16.1695%;\\\"\\u003e\\n \\u003cp\\u003e\\u0026nbsp;Clinically incomplete\\u003c/p\\u003e\\n \\u003c/td\\u003e\\n \\u003ctd valign=\\\"top\\\" style=\\\"width: 6.894%;\\\"\\u003e\\n \\u003cp\\u003e19/31 (61.3%)\\u003c/p\\u003e\\n \\u003c/td\\u003e\\n \\u003ctd valign=\\\"top\\\" style=\\\"width: 12.9105%;\\\"\\u003e\\n \\u003cp\\u003e209/500 (41.8%) responses deemed not useful; 8/20 (40.0%) poor guideline adherence; 42/100 (42.0%) AMR mechanism errors\\u003c/p\\u003e\\n \\u003c/td\\u003e\\n \\u003ctd valign=\\\"top\\\" style=\\\"width: 7.7714%;\\\"\\u003e\\n \\u003cp\\u003eModerate\\u003c/p\\u003e\\n \\u003c/td\\u003e\\n \\u003ctd valign=\\\"top\\\" style=\\\"width: 10.0276%;\\\"\\u003e\\n \\u003cp\\u003eYes\\u003c/p\\u003e\\n \\u003c/td\\u003e\\n \\u003ctd valign=\\\"top\\\" style=\\\"width: 11.4064%;\\\"\\u003e\\n \\u003cp\\u003eIsolation/IPC, guideline adherence, antimicrobial choice, citations, AMR interpretation, diagnostics, dosing\\u003c/p\\u003e\\n \\u003c/td\\u003e\\n \\u003ctd valign=\\\"top\\\" style=\\\"width: 10.0276%;\\\"\\u003e\\n \\u003cp\\u003eChain-of-thought prompting, prompt engineering, external validation\\u003c/p\\u003e\\n \\u003c/td\\u003e\\n \\u003ctd valign=\\\"top\\\" style=\\\"width: 11.7824%;\\\"\\u003e\\n \\u003cp\\u003eAbosi 2024, Kufel 2024\\u003c/p\\u003e\\n \\u003c/td\\u003e\\n \\u003c/tr\\u003e\\n \\u003ctr\\u003e\\n \\u003ctd valign=\\\"top\\\" style=\\\"width: 16.1695%;\\\"\\u003e\\n \\u003cp\\u003eConstraint-induced\\u003c/p\\u003e\\n \\u003c/td\\u003e\\n \\u003ctd valign=\\\"top\\\" style=\\\"width: 6.894%;\\\"\\u003e\\n \\u003cp\\u003e9/31 (29.0%)\\u003c/p\\u003e\\n \\u003c/td\\u003e\\n \\u003ctd valign=\\\"top\\\" style=\\\"width: 12.9105%;\\\"\\u003e\\n \\u003cp\\u003e31/393 (7.9%) unanswered queries; 11/17 (64.7%) errors from missing chart information\\u003c/p\\u003e\\n \\u003c/td\\u003e\\n \\u003ctd valign=\\\"top\\\" style=\\\"width: 7.7714%;\\\"\\u003e\\n \\u003cp\\u003eLow\\u003c/p\\u003e\\n \\u003c/td\\u003e\\n \\u003ctd valign=\\\"top\\\" style=\\\"width: 10.0276%;\\\"\\u003e\\n \\u003cp\\u003eYes\\u003c/p\\u003e\\n \\u003c/td\\u003e\\n \\u003ctd valign=\\\"top\\\" style=\\\"width: 11.4064%;\\\"\\u003e\\n \\u003cp\\u003eDiagnostics, extraction\\u003c/p\\u003e\\n \\u003c/td\\u003e\\n \\u003ctd valign=\\\"top\\\" style=\\\"width: 10.0276%;\\\"\\u003e\\n \\u003cp\\u003eRAG, expanded context windows, structured input templates\\u003c/p\\u003e\\n \\u003c/td\\u003e\\n \\u003ctd valign=\\\"top\\\" style=\\\"width: 11.7824%;\\\"\\u003e\\n \\u003cp\\u003eLorenzoni 2025, Rodriguez-Nava 2025\\u003c/p\\u003e\\n \\u003c/td\\u003e\\n \\u003c/tr\\u003e\\n \\u003ctr\\u003e\\n \\u003ctd valign=\\\"top\\\" style=\\\"width: 16.1695%;\\\"\\u003e\\n \\u003cp\\u003eInternal contradictions / instability\\u003c/p\\u003e\\n \\u003c/td\\u003e\\n \\u003ctd valign=\\\"top\\\" style=\\\"width: 6.894%;\\\"\\u003e\\n \\u003cp\\u003e2/31 (6.5%)\\u003c/p\\u003e\\n \\u003c/td\\u003e\\n \\u003ctd valign=\\\"top\\\" style=\\\"width: 12.9105%;\\\"\\u003e\\n \\u003cp\\u003enot reported\\u003c/p\\u003e\\n \\u003c/td\\u003e\\n \\u003ctd valign=\\\"top\\\" style=\\\"width: 7.7714%;\\\"\\u003e\\n \\u003cp\\u003eModerate\\u003c/p\\u003e\\n \\u003c/td\\u003e\\n \\u003ctd valign=\\\"top\\\" style=\\\"width: 10.0276%;\\\"\\u003e\\n \\u003cp\\u003eUnclear\\u003c/p\\u003e\\n \\u003c/td\\u003e\\n \\u003ctd valign=\\\"top\\\" style=\\\"width: 11.4064%;\\\"\\u003e\\n \\u003cp\\u003eGuideline adherence, antimicrobial choice\\u003c/p\\u003e\\n \\u003c/td\\u003e\\n \\u003ctd valign=\\\"top\\\" style=\\\"width: 10.0276%;\\\"\\u003e\\n \\u003cp\\u003eVersion pinning, multiple runs with consensus\\u003c/p\\u003e\\n \\u003c/td\\u003e\\n \\u003ctd valign=\\\"top\\\" style=\\\"width: 11.7824%;\\\"\\u003e\\n \\u003cp\\u003eCakir 2023, Borgonovo 2025\\u003c/p\\u003e\\n \\u003c/td\\u003e\\n \\u003c/tr\\u003e\\n \\u003ctr\\u003e\\n \\u003ctd valign=\\\"top\\\" style=\\\"width: 16.1695%;\\\"\\u003e\\n \\u003cp\\u003eContext-dependent failures\\u003c/p\\u003e\\n \\u003c/td\\u003e\\n \\u003ctd valign=\\\"top\\\" style=\\\"width: 6.894%;\\\"\\u003e\\n \\u003cp\\u003e3/31 (9.7%)\\u003c/p\\u003e\\n \\u003c/td\\u003e\\n \\u003ctd valign=\\\"top\\\" style=\\\"width: 12.9105%;\\\"\\u003e\\n \\u003cp\\u003eSpecificity improved from 35% to 75% when complete chart provided\\u003c/p\\u003e\\n \\u003c/td\\u003e\\n \\u003ctd valign=\\\"top\\\" style=\\\"width: 7.7714%;\\\"\\u003e\\n \\u003cp\\u003eModerate\\u003c/p\\u003e\\n \\u003c/td\\u003e\\n \\u003ctd valign=\\\"top\\\" style=\\\"width: 10.0276%;\\\"\\u003e\\n \\u003cp\\u003eYes\\u003c/p\\u003e\\n \\u003c/td\\u003e\\n \\u003ctd valign=\\\"top\\\" style=\\\"width: 11.4064%;\\\"\\u003e\\n \\u003cp\\u003eDiagnostics, surveillance\\u003c/p\\u003e\\n \\u003c/td\\u003e\\n \\u003ctd valign=\\\"top\\\" style=\\\"width: 10.0276%;\\\"\\u003e\\n \\u003cp\\u003eRAG, full chart access, external data integration\\u003c/p\\u003e\\n \\u003c/td\\u003e\\n \\u003ctd valign=\\\"top\\\" style=\\\"width: 11.7824%;\\\"\\u003e\\n \\u003cp\\u003eRodriguez-Nava 2025, Wiemken 2024\\u003c/p\\u003e\\n \\u003c/td\\u003e\\n \\u003c/tr\\u003e\\n \\u003ctr\\u003e\\n \\u003ctd valign=\\\"top\\\" style=\\\"width: 16.1695%;\\\"\\u003e\\n \\u003cp\\u003eAny safety event\\u003c/p\\u003e\\n \\u003c/td\\u003e\\n \\u003ctd valign=\\\"top\\\" style=\\\"width: 6.894%;\\\"\\u003e\\n \\u003cp\\u003e\\u003cstrong\\u003e28/31 (90.3%)\\u003c/strong\\u003e\\u003c/p\\u003e\\n \\u003c/td\\u003e\\n \\u003ctd valign=\\\"top\\\" style=\\\"width: 12.9105%;\\\"\\u003e\\n \\u003cp\\u003e\\u0026mdash;\\u003c/p\\u003e\\n \\u003c/td\\u003e\\n \\u003ctd valign=\\\"top\\\" style=\\\"width: 7.7714%;\\\"\\u003e\\n \\u003cp\\u003e\\u0026mdash;\\u003c/p\\u003e\\n \\u003c/td\\u003e\\n \\u003ctd valign=\\\"top\\\" style=\\\"width: 10.0276%;\\\"\\u003e\\n \\u003cp\\u003e\\u0026mdash;\\u003c/p\\u003e\\n \\u003c/td\\u003e\\n \\u003ctd valign=\\\"top\\\" style=\\\"width: 11.4064%;\\\"\\u003e\\n \\u003cp\\u003e\\u0026mdash;\\u003c/p\\u003e\\n \\u003c/td\\u003e\\n \\u003ctd valign=\\\"top\\\" style=\\\"width: 10.0276%;\\\"\\u003e\\n \\u003cp\\u003e\\u0026mdash;\\u003c/p\\u003e\\n \\u003c/td\\u003e\\n \\u003ctd valign=\\\"top\\\" style=\\\"width: 11.7824%;\\\"\\u003e\\n \\u003cp\\u003e\\u0026mdash;\\u003c/p\\u003e\\n \\u003c/td\\u003e\\n \\u003c/tr\\u003e\\n \\u003c/tbody\\u003e\\n\\u003c/table\\u003e\\n\\u003cp\\u003e\\u003cstrong\\u003eFootnotes:\\u003c/strong\\u003e\\u003c/p\\u003e\\n\\u003col\\u003e\\n \\u003cli\\u003e\\u003cstrong\\u003eDefinitions:\\u003c/strong\\u003e Unsafe recommendations = clearly incorrect/contraindicated management; Harmful/inadequate = omissions (e.g., source control) or dosing errors with potential for patient harm; Hallucinations = fabricated citations, contradictory statements, or artificial clinical details; Clinically incomplete = missing key clinical elements despite being prompted; Constraint-induced = refusals, guardrail blocks, verbosity, token limits, or language barriers preventing complete responses; Internal contradictions = intra-session inconsistencies in recommendations; Context-dependent = errors mitigated by providing full chart access or external knowledge.\\u003c/li\\u003e\\n \\u003cli\\u003ePercentages reflect unique studies per category. Studies may appear in multiple categories if they reported multiple safety issue types.\\u003c/li\\u003e\\n \\u003cli\\u003e\\u0026quot;Not reported\\u0026quot; indicates no numeric denominator was available in the source studies despite qualitative safety signals being documented.\\u003c/li\\u003e\\n \\u003cli\\u003e\\u0026quot;Context-dependent\\u0026quot; indicates issues that improved with RAG, full chart access, or expert oversight in reported studies.\\u003c/li\\u003e\\n \\u003cli\\u003eCommon mitigations listed represent strategies that were tested and showed some benefit across studies, though not all were formally quantified.\\u003c/li\\u003e\\n\\u003c/ol\\u003e\\n\\u003cp\\u003e\\u003cstrong\\u003eTable 2. \\u003c/strong\\u003eStudy Limitations and Clinical Applicability Constraints\\u003c/p\\u003e\\n\\u003ctable border=\\\"1\\\" cellspacing=\\\"0\\\" cellpadding=\\\"0\\\" width=\\\"794\\\"\\u003e\\n \\u003ctbody\\u003e\\n \\u003ctr\\u003e\\n \\u003ctd valign=\\\"top\\\" style=\\\"width: 214px;\\\"\\u003e\\n \\u003cp\\u003e\\u003cstrong\\u003eLimitation Category\\u003c/strong\\u003e\\u003c/p\\u003e\\n \\u003c/td\\u003e\\n \\u003ctd valign=\\\"top\\\" style=\\\"width: 70px;\\\"\\u003e\\n \\u003cp\\u003e\\u003cstrong\\u003eStudies Affected (n/31; %)\\u003c/strong\\u003e\\u003c/p\\u003e\\n \\u003c/td\\u003e\\n \\u003ctd valign=\\\"top\\\" style=\\\"width: 246px;\\\"\\u003e\\n \\u003cp\\u003e\\u003cstrong\\u003eDescription and Examples\\u003c/strong\\u003e\\u003c/p\\u003e\\n \\u003c/td\\u003e\\n \\u003ctd valign=\\\"top\\\" style=\\\"width: 264px;\\\"\\u003e\\n \\u003cp\\u003e\\u003cstrong\\u003eImpact on Clinical Applicability\\u003c/strong\\u003e\\u003c/p\\u003e\\n \\u003c/td\\u003e\\n \\u003c/tr\\u003e\\n \\u003ctr\\u003e\\n \\u003ctd colspan=\\\"3\\\" valign=\\\"top\\\" style=\\\"width: 530px;\\\"\\u003e\\n \\u003cp\\u003e\\u003cstrong\\u003eStudy Design and Validation\\u003c/strong\\u003e\\u003c/p\\u003e\\n \\u003c/td\\u003e\\n \\u003ctd valign=\\\"top\\\" style=\\\"width: 264px;\\\"\\u003e\\u003cbr\\u003e\\u003c/td\\u003e\\n \\u003c/tr\\u003e\\n \\u003ctr\\u003e\\n \\u003ctd valign=\\\"top\\\" style=\\\"width: 214px;\\\"\\u003e\\n \\u003cp\\u003eVignette/synthetic cases\\u003c/p\\u003e\\n \\u003c/td\\u003e\\n \\u003ctd valign=\\\"top\\\" style=\\\"width: 70px;\\\"\\u003e\\n \\u003cp\\u003e21/31 (67.74%)\\u003c/p\\u003e\\n \\u003c/td\\u003e\\n \\u003ctd valign=\\\"top\\\" style=\\\"width: 246px;\\\"\\u003e\\n \\u003cp\\u003eSimulated rather than real clinical scenarios; may not capture complexity\\u003c/p\\u003e\\n \\u003c/td\\u003e\\n \\u003ctd valign=\\\"top\\\" style=\\\"width: 264px;\\\"\\u003e\\n \\u003cp\\u003eMay overestimate performance in actual practice\\u003c/p\\u003e\\n \\u003c/td\\u003e\\n \\u003c/tr\\u003e\\n \\u003ctr\\u003e\\n \\u003ctd valign=\\\"top\\\" style=\\\"width: 214px;\\\"\\u003e\\n \\u003cp\\u003eSmall sample size\\u003c/p\\u003e\\n \\u003c/td\\u003e\\n \\u003ctd valign=\\\"top\\\" style=\\\"width: 70px;\\\"\\u003e\\n \\u003cp\\u003e11/31 (35%)\\u003c/p\\u003e\\n \\u003c/td\\u003e\\n \\u003ctd valign=\\\"top\\\" style=\\\"width: 246px;\\\"\\u003e\\n \\u003cp\\u003e\\u0026lt;50 cases evaluated (range: 7-44)\\u003c/p\\u003e\\n \\u003c/td\\u003e\\n \\u003ctd valign=\\\"top\\\" style=\\\"width: 264px;\\\"\\u003e\\n \\u003cp\\u003eInsufficient power; wide confidence intervals\\u003c/p\\u003e\\n \\u003c/td\\u003e\\n \\u003c/tr\\u003e\\n \\u003ctr\\u003e\\n \\u003ctd valign=\\\"top\\\" style=\\\"width: 214px;\\\"\\u003e\\n \\u003cp\\u003eSingle-center studies\\u003c/p\\u003e\\n \\u003c/td\\u003e\\n \\u003ctd valign=\\\"top\\\" style=\\\"width: 70px;\\\"\\u003e\\n \\u003cp\\u003e8/31 (26%)\\u003c/p\\u003e\\n \\u003c/td\\u003e\\n \\u003ctd valign=\\\"top\\\" style=\\\"width: 246px;\\\"\\u003e\\n \\u003cp\\u003eLimited to one institution\\u0026apos;s practices/population\\u003c/p\\u003e\\n \\u003c/td\\u003e\\n \\u003ctd valign=\\\"top\\\" style=\\\"width: 264px;\\\"\\u003e\\n \\u003cp\\u003eGeneralizability uncertain across healthcare systems\\u003c/p\\u003e\\n \\u003c/td\\u003e\\n \\u003c/tr\\u003e\\n \\u003ctr\\u003e\\n \\u003ctd colspan=\\\"3\\\" valign=\\\"top\\\" style=\\\"width: 530px;\\\"\\u003e\\n \\u003cp\\u003e\\u003cstrong\\u003eData and Context Constraints\\u003c/strong\\u003e\\u003c/p\\u003e\\n \\u003c/td\\u003e\\n \\u003ctd valign=\\\"top\\\" style=\\\"width: 264px;\\\"\\u003e\\u003cbr\\u003e\\u003c/td\\u003e\\n \\u003c/tr\\u003e\\n \\u003ctr\\u003e\\n \\u003ctd valign=\\\"top\\\" style=\\\"width: 214px;\\\"\\u003e\\n \\u003cp\\u003eLanguage limitations\\u003c/p\\u003e\\n \\u003c/td\\u003e\\n \\u003ctd valign=\\\"top\\\" style=\\\"width: 70px;\\\"\\u003e\\n \\u003cp\\u003e3/31 (10%)\\u003c/p\\u003e\\n \\u003c/td\\u003e\\n \\u003ctd valign=\\\"top\\\" style=\\\"width: 246px;\\\"\\u003e\\n \\u003cp\\u003eClinical documentation in non-English (Italian, Turkish); chatbots English-only\\u003c/p\\u003e\\n \\u003c/td\\u003e\\n \\u003ctd valign=\\\"top\\\" style=\\\"width: 264px;\\\"\\u003e\\n \\u003cp\\u003eReduced accuracy for non-English contexts; limits global applicability\\u003c/p\\u003e\\n \\u003c/td\\u003e\\n \\u003c/tr\\u003e\\n \\u003ctr\\u003e\\n \\u003ctd valign=\\\"top\\\" style=\\\"width: 214px;\\\"\\u003e\\n \\u003cp\\u003eToken/context limits\\u003c/p\\u003e\\n \\u003c/td\\u003e\\n \\u003ctd valign=\\\"top\\\" style=\\\"width: 70px;\\\"\\u003e\\n \\u003cp\\u003e5/31 (16%)\\u003c/p\\u003e\\n \\u003c/td\\u003e\\n \\u003ctd valign=\\\"top\\\" style=\\\"width: 246px;\\\"\\u003e\\n \\u003cp\\u003eUnable to process complete medical records; limited to recent 2 progress notes\\u003c/p\\u003e\\n \\u003c/td\\u003e\\n \\u003ctd valign=\\\"top\\\" style=\\\"width: 264px;\\\"\\u003e\\n \\u003cp\\u003eCritical information missed; specificity dropped 35%\\u0026rarr;75% when full chart provided\\u003c/p\\u003e\\n \\u003c/td\\u003e\\n \\u003c/tr\\u003e\\n \\u003ctr\\u003e\\n \\u003ctd valign=\\\"top\\\" style=\\\"width: 214px;\\\"\\u003e\\n \\u003cp\\u003eKnowledge cutoff issues\\u003c/p\\u003e\\n \\u003c/td\\u003e\\n \\u003ctd valign=\\\"top\\\" style=\\\"width: 70px;\\\"\\u003e\\n \\u003cp\\u003e4/31 (13%)\\u003c/p\\u003e\\n \\u003c/td\\u003e\\n \\u003ctd valign=\\\"top\\\" style=\\\"width: 246px;\\\"\\u003e\\n \\u003cp\\u003eTraining data ended 2021; inconsistent with latest guidelines (e.g., 2023 HBV consensus)\\u003c/p\\u003e\\n \\u003c/td\\u003e\\n \\u003ctd valign=\\\"top\\\" style=\\\"width: 264px;\\\"\\u003e\\n \\u003cp\\u003eOutdated recommendations; missed recent evidence\\u003c/p\\u003e\\n \\u003c/td\\u003e\\n \\u003c/tr\\u003e\\n \\u003ctr\\u003e\\n \\u003ctd valign=\\\"top\\\" style=\\\"width: 214px;\\\"\\u003e\\n \\u003cp\\u003eMissing local context\\u003c/p\\u003e\\n \\u003c/td\\u003e\\n \\u003ctd valign=\\\"top\\\" style=\\\"width: 70px;\\\"\\u003e\\n \\u003cp\\u003e6/31 (19%)\\u003c/p\\u003e\\n \\u003c/td\\u003e\\n \\u003ctd valign=\\\"top\\\" style=\\\"width: 246px;\\\"\\u003e\\n \\u003cp\\u003eNo integration of local antibiograms, formularies, or resistance patterns\\u003c/p\\u003e\\n \\u003c/td\\u003e\\n \\u003ctd valign=\\\"top\\\" style=\\\"width: 264px;\\\"\\u003e\\n \\u003cp\\u003eRecommendations may be inappropriate/unavailable regionally\\u003c/p\\u003e\\n \\u003c/td\\u003e\\n \\u003c/tr\\u003e\\n \\u003ctr\\u003e\\n \\u003ctd colspan=\\\"2\\\" valign=\\\"top\\\" style=\\\"width: 284px;\\\"\\u003e\\n \\u003cp\\u003e\\u003cstrong\\u003eGovernance and Safety\\u003c/strong\\u003e\\u003c/p\\u003e\\n \\u003c/td\\u003e\\n \\u003ctd valign=\\\"top\\\" style=\\\"width: 246px;\\\"\\u003e\\u003cbr\\u003e\\u003c/td\\u003e\\n \\u003ctd valign=\\\"top\\\" style=\\\"width: 264px;\\\"\\u003e\\u003cbr\\u003e\\u003c/td\\u003e\\n \\u003c/tr\\u003e\\n \\u003ctr\\u003e\\n \\u003ctd valign=\\\"top\\\" style=\\\"width: 214px;\\\"\\u003e\\n \\u003cp\\u003eNo PHI protections\\u003c/p\\u003e\\n \\u003c/td\\u003e\\n \\u003ctd valign=\\\"top\\\" style=\\\"width: 70px;\\\"\\u003e\\n \\u003cp\\u003e12/31 (39%)\\u003c/p\\u003e\\n \\u003c/td\\u003e\\n \\u003ctd valign=\\\"top\\\" style=\\\"width: 246px;\\\"\\u003e\\n \\u003cp\\u003ePublic web interfaces; no HIPAA-compliant/secure instances\\u003c/p\\u003e\\n \\u003c/td\\u003e\\n \\u003ctd valign=\\\"top\\\" style=\\\"width: 264px;\\\"\\u003e\\n \\u003cp\\u003eCannot be used with real patient data in most settings\\u003c/p\\u003e\\n \\u003c/td\\u003e\\n \\u003c/tr\\u003e\\n \\u003ctr\\u003e\\n \\u003ctd valign=\\\"top\\\" style=\\\"width: 214px;\\\"\\u003e\\n \\u003cp\\u003eIRB not required/waived\\u003c/p\\u003e\\n \\u003c/td\\u003e\\n \\u003ctd valign=\\\"top\\\" style=\\\"width: 70px;\\\"\\u003e\\n \\u003cp\\u003e8/31 (26%)\\u003c/p\\u003e\\n \\u003c/td\\u003e\\n \\u003ctd valign=\\\"top\\\" style=\\\"width: 246px;\\\"\\u003e\\n \\u003cp\\u003eClassified as non-human subjects research or quality improvement\\u003c/p\\u003e\\n \\u003c/td\\u003e\\n \\u003ctd valign=\\\"top\\\" style=\\\"width: 264px;\\\"\\u003e\\n \\u003cp\\u003eEthical oversight gaps for clinical deployment\\u003c/p\\u003e\\n \\u003c/td\\u003e\\n \\u003c/tr\\u003e\\n \\u003ctr\\u003e\\n \\u003ctd valign=\\\"top\\\" style=\\\"width: 214px;\\\"\\u003e\\n \\u003cp\\u003eLack of regulatory alignment\\u003c/p\\u003e\\n \\u003c/td\\u003e\\n \\u003ctd valign=\\\"top\\\" style=\\\"width: 70px;\\\"\\u003e\\n \\u003cp\\u003e23/31 (74%)\\u003c/p\\u003e\\n \\u003c/td\\u003e\\n \\u003ctd valign=\\\"top\\\" style=\\\"width: 246px;\\\"\\u003e\\n \\u003cp\\u003eNo FDA/CE marking or clinical validation pathway\\u003c/p\\u003e\\n \\u003c/td\\u003e\\n \\u003ctd valign=\\\"top\\\" style=\\\"width: 264px;\\\"\\u003e\\n \\u003cp\\u003eUnclear regulatory status for clinical use\\u003c/p\\u003e\\n \\u003c/td\\u003e\\n \\u003c/tr\\u003e\\n \\u003ctr\\u003e\\n \\u003ctd colspan=\\\"3\\\" valign=\\\"top\\\" style=\\\"width: 530px;\\\"\\u003e\\n \\u003cp\\u003e\\u003cstrong\\u003eImplementation and Usability\\u003c/strong\\u003e\\u003c/p\\u003e\\n \\u003c/td\\u003e\\n \\u003ctd valign=\\\"top\\\" style=\\\"width: 264px;\\\"\\u003e\\u003cbr\\u003e\\u003c/td\\u003e\\n \\u003c/tr\\u003e\\n \\u003ctr\\u003e\\n \\u003ctd valign=\\\"top\\\" style=\\\"width: 214px;\\\"\\u003e\\n \\u003cp\\u003eVerbosity/poor readability\\u003c/p\\u003e\\n \\u003c/td\\u003e\\n \\u003ctd valign=\\\"top\\\" style=\\\"width: 70px;\\\"\\u003e\\n \\u003cp\\u003e5/31 (16%)\\u003c/p\\u003e\\n \\u003c/td\\u003e\\n \\u003ctd valign=\\\"top\\\" style=\\\"width: 246px;\\\"\\u003e\\n \\u003cp\\u003eResponses 200-400+ words; reading level \\u0026gt;10th grade vs recommended 6th grade\\u003c/p\\u003e\\n \\u003c/td\\u003e\\n \\u003ctd valign=\\\"top\\\" style=\\\"width: 264px;\\\"\\u003e\\n \\u003cp\\u003eUnusable in time-critical settings; incomprehensible to patients\\u003c/p\\u003e\\n \\u003c/td\\u003e\\n \\u003c/tr\\u003e\\n \\u003ctr\\u003e\\n \\u003ctd valign=\\\"top\\\" style=\\\"width: 214px;\\\"\\u003e\\n \\u003cp\\u003eInconsistent reproducibility\\u003c/p\\u003e\\n \\u003c/td\\u003e\\n \\u003ctd valign=\\\"top\\\" style=\\\"width: 70px;\\\"\\u003e\\n \\u003cp\\u003e4/31 (13%)\\u003c/p\\u003e\\n \\u003c/td\\u003e\\n \\u003ctd valign=\\\"top\\\" style=\\\"width: 246px;\\\"\\u003e\\n \\u003cp\\u003eDifferent answers when asked repeatedly; within-session contradictions\\u003c/p\\u003e\\n \\u003c/td\\u003e\\n \\u003ctd valign=\\\"top\\\" style=\\\"width: 264px;\\\"\\u003e\\n \\u003cp\\u003eUnreliable for clinical decision-making\\u003c/p\\u003e\\n \\u003c/td\\u003e\\n \\u003c/tr\\u003e\\n \\u003ctr\\u003e\\n \\u003ctd valign=\\\"top\\\" style=\\\"width: 214px;\\\"\\u003e\\n \\u003cp\\u003eConstraint-induced failures\\u003c/p\\u003e\\n \\u003c/td\\u003e\\n \\u003ctd valign=\\\"top\\\" style=\\\"width: 70px;\\\"\\u003e\\n \\u003cp\\u003e9/31 (29%)\\u003c/p\\u003e\\n \\u003c/td\\u003e\\n \\u003ctd valign=\\\"top\\\" style=\\\"width: 246px;\\\"\\u003e\\n \\u003cp\\u003eGuardrail refusals, unanswered queries (8% of interactions), or excessive caveats\\u003c/p\\u003e\\n \\u003c/td\\u003e\\n \\u003ctd valign=\\\"top\\\" style=\\\"width: 264px;\\\"\\u003e\\n \\u003cp\\u003eIncomplete clinical utility; frustrating user experience\\u003c/p\\u003e\\n \\u003c/td\\u003e\\n \\u003c/tr\\u003e\\n \\u003ctr\\u003e\\n \\u003ctd valign=\\\"top\\\" style=\\\"width: 214px;\\\"\\u003e\\n \\u003cp\\u003eNo uncertainty quantification\\u003c/p\\u003e\\n \\u003c/td\\u003e\\n \\u003ctd valign=\\\"top\\\" style=\\\"width: 70px;\\\"\\u003e\\n \\u003cp\\u003e6/31 (19%)\\u003c/p\\u003e\\n \\u003c/td\\u003e\\n \\u003ctd valign=\\\"top\\\" style=\\\"width: 246px;\\\"\\u003e\\n \\u003cp\\u003eFailed to acknowledge limitations or express appropriate uncertainty\\u003c/p\\u003e\\n \\u003c/td\\u003e\\n \\u003ctd valign=\\\"top\\\" style=\\\"width: 264px;\\\"\\u003e\\n \\u003cp\\u003eFalse confidence; safety risk\\u003c/p\\u003e\\n \\u003c/td\\u003e\\n \\u003c/tr\\u003e\\n \\u003ctr\\u003e\\n \\u003ctd colspan=\\\"2\\\" valign=\\\"top\\\" style=\\\"width: 284px;\\\"\\u003e\\n \\u003cp\\u003e\\u003cstrong\\u003eClinical Preparedness Gaps\\u003c/strong\\u003e\\u003c/p\\u003e\\n \\u003c/td\\u003e\\n \\u003ctd valign=\\\"top\\\" style=\\\"width: 246px;\\\"\\u003e\\u003cbr\\u003e\\u003c/td\\u003e\\n \\u003ctd valign=\\\"top\\\" style=\\\"width: 264px;\\\"\\u003e\\u003cbr\\u003e\\u003c/td\\u003e\\n \\u003c/tr\\u003e\\n \\u003ctr\\u003e\\n \\u003ctd valign=\\\"top\\\" style=\\\"width: 214px;\\\"\\u003e\\n \\u003cp\\u003eNo real-time EHR integration\\u003c/p\\u003e\\n \\u003c/td\\u003e\\n \\u003ctd valign=\\\"top\\\" style=\\\"width: 70px;\\\"\\u003e\\n \\u003cp\\u003e27/31 (87%)\\u003c/p\\u003e\\n \\u003c/td\\u003e\\n \\u003ctd valign=\\\"top\\\" style=\\\"width: 246px;\\\"\\u003e\\n \\u003cp\\u003eManual data entry required; cannot access lab/imaging/notes automatically\\u003c/p\\u003e\\n \\u003c/td\\u003e\\n \\u003ctd valign=\\\"top\\\" style=\\\"width: 264px;\\\"\\u003e\\n \\u003cp\\u003eImpractical for clinical workflow; introduces transcription errors\\u003c/p\\u003e\\n \\u003c/td\\u003e\\n \\u003c/tr\\u003e\\n \\u003ctr\\u003e\\n \\u003ctd valign=\\\"top\\\" style=\\\"width: 214px;\\\"\\u003e\\n \\u003cp\\u003eNo human-in-the-loop validation\\u003c/p\\u003e\\n \\u003c/td\\u003e\\n \\u003ctd valign=\\\"top\\\" style=\\\"width: 70px;\\\"\\u003e\\n \\u003cp\\u003e19/31 (61%)\\u003c/p\\u003e\\n \\u003c/td\\u003e\\n \\u003ctd valign=\\\"top\\\" style=\\\"width: 246px;\\\"\\u003e\\n \\u003cp\\u003eNo expert review mechanism before recommendations acted upon\\u003c/p\\u003e\\n \\u003c/td\\u003e\\n \\u003ctd valign=\\\"top\\\" style=\\\"width: 264px;\\\"\\u003e\\n \\u003cp\\u003eUnsafe for autonomous use\\u003c/p\\u003e\\n \\u003c/td\\u003e\\n \\u003c/tr\\u003e\\n \\u003ctr\\u003e\\n \\u003ctd valign=\\\"top\\\" style=\\\"width: 214px;\\\"\\u003e\\n \\u003cp\\u003eInsufficient process measures\\u003c/p\\u003e\\n \\u003c/td\\u003e\\n \\u003ctd valign=\\\"top\\\" style=\\\"width: 70px;\\\"\\u003e\\n \\u003cp\\u003e21/31 (68%)\\u003c/p\\u003e\\n \\u003c/td\\u003e\\n \\u003ctd valign=\\\"top\\\" style=\\\"width: 246px;\\\"\\u003e\\n \\u003cp\\u003eIntervention fidelity not reported; unclear what drove observed effects\\u003c/p\\u003e\\n \\u003c/td\\u003e\\n \\u003ctd valign=\\\"top\\\" style=\\\"width: 264px;\\\"\\u003e\\n \\u003cp\\u003eCannot determine which implementation strategies effective\\u003c/p\\u003e\\n \\u003c/td\\u003e\\n \\u003c/tr\\u003e\\n \\u003c/tbody\\u003e\\n\\u003c/table\\u003e\\n\\u003cp\\u003e\\u003cstrong\\u003eFootnotes:\\u003c/strong\\u003e Percentages represent unique studies reporting each limitation category. Categories are not mutually exclusive. EHR=electronic health record; HBV=hepatitis B virus; IRB=institutional review board; PHI=protected health information.\\u003c/p\\u003e\"}],\"fulltextSource\":\"\",\"fullText\":\"\",\"funders\":[{\"identity\":\"94079679-ddb6-4d4c-8f5a-30764ddc59c9\",\"identifier\":\"10.13039/100006108\",\"name\":\"National Center for Advancing Translational Sciences\",\"awardNumber\":\"UL1TR004419 \",\"order_by\":0},{\"identity\":\"766a69a6-1c2e-4248-b9f5-72fdb7ffca6c\",\"identifier\":\"10.13039/100000002\",\"name\":\"National Institutes of Health\",\"awardNumber\":\"S10OD026880\",\"order_by\":1},{\"identity\":\"3c8cced2-ed53-4dff-addb-1d269e418fda\",\"identifier\":\"10.13039/100000002\",\"name\":\"National Institutes of Health\",\"awardNumber\":\"S10OD030463\",\"order_by\":2}],\"hasAdminPriorityOnWorkflow\":false,\"hasManuscriptDocX\":true,\"hasOptedInToPreprint\":true,\"hasPassedJournalQc\":\"\",\"hasAnyPriority\":true,\"hideJournal\":true,\"highlight\":\"\",\"institution\":\"Icahn School of Medicine at Mount Sinai\",\"isAcceptedByJournal\":false,\"isAuthorSuppliedPdf\":false,\"isDeskRejected\":\"\",\"isHiddenFromSearch\":false,\"isInQc\":false,\"isInWorkflow\":false,\"isPdf\":false,\"isPdfUpToDate\":true,\"isWithdrawnOrRetracted\":false,\"journal\":{\"display\":true,\"email\":\"info@researchsquare.com\",\"identity\":\"researchsquare\",\"isNatureJournal\":false,\"hasQc\":true,\"allowDirectSubmit\":true,\"externalIdentity\":\"\",\"sideBox\":\"\",\"snPcode\":\"\",\"submissionUrl\":\"/submission\",\"title\":\"Research Square\",\"twitterHandle\":\"researchsquare\",\"acdcEnabled\":true,\"dfaEnabled\":false,\"editorialSystem\":\"\",\"reportingPortfolio\":\"\",\"inReviewEnabled\":false,\"inReviewRevisionsEnabled\":true},\"keywords\":\"Infectious diseases, Large language models, Clinical decision support, Antimicrobial stewardship, Hallucinations (AI), Patient safety, Retrieval-augmented generation, Bias and fairness\",\"lastPublishedDoi\":\"10.21203/rs.3.rs-8901882/v1\",\"lastPublishedDoiUrl\":\"https://doi.org/10.21203/rs.3.rs-8901882/v1\",\"license\":{\"name\":\"CC BY 4.0\",\"url\":\"https://creativecommons.org/licenses/by/4.0/\"},\"manuscriptAbstract\":\"\\u003ch2\\u003eBackground\\u003c/h2\\u003e \\u003cp\\u003eClinical reasoning in infectious diseases relies on validated evidence. LLMs are being introduced into diagnosis, antimicrobial stewardship, and guideline interpretation before their safety and reliability are established.\\u003c/p\\u003e\\u003ch2\\u003eMethods\\u003c/h2\\u003e \\u003cp\\u003eThis review, registered in PROSPERO (CRD420251155354), evaluated studies using GPT, Claude, Gemini, and retrieval-augmented or agentic systems for infectious disease decision-making. PubMed, CENTRAL, Scopus, and Web of Science were searched from January 2018 to September 2025. Two reviewers screened and extracted data. Risk of bias was assessed with QUADAS-AI.\\u003c/p\\u003e\\u003ch2\\u003eFindings:\\u003c/h2\\u003e \\u003cp\\u003eThirty-one studies met inclusion criteria. Most were cross-sectional (61%) and vignette-based (68%). Only 32% used real clinical data; 23% had low risk of bias. Safety issues were reported in 90% of studies: incomplete responses (61%), unsafe advice (23\\u0026ndash;32%), and fabricated content (32%). In antimicrobial stewardship, agreement with infectious-disease specialists was ~\\u0026thinsp;50%. Diagnostic sensitivity for structured infections was 80\\u0026ndash;100%. Retrieval-augmented systems increased specificity from 35% to 75% and reduced hallucinations. Proprietary models outperformed open-source models but did not reach expert accuracy.\\u003c/p\\u003e\\u003ch2\\u003eInterpretation:\\u003c/h2\\u003e \\u003cp\\u003eLLMs perform well in defined diagnostic tasks but remain unreliable for autonomous clinical use. High error rates, inconsistent reasoning, and fabricated content require expert oversight and external validation before deployment.\\u003c/p\\u003e\",\"manuscriptTitle\":\"Large Language Models in Infectious Diseases: A Systemic Review\",\"msid\":\"\",\"msnumber\":\"\",\"nonDraftVersions\":[{\"code\":1,\"date\":\"2026-02-18 07:27:53\",\"doi\":\"10.21203/rs.3.rs-8901882/v1\",\"editorialEvents\":[{\"type\":\"communityComments\",\"content\":0}],\"status\":\"published\",\"journal\":{\"display\":true,\"email\":\"info@researchsquare.com\",\"identity\":\"researchsquare\",\"isNatureJournal\":false,\"hasQc\":true,\"allowDirectSubmit\":true,\"externalIdentity\":\"\",\"sideBox\":\"\",\"snPcode\":\"\",\"submissionUrl\":\"/submission\",\"title\":\"Research Square\",\"twitterHandle\":\"researchsquare\",\"acdcEnabled\":true,\"dfaEnabled\":false,\"editorialSystem\":\"\",\"reportingPortfolio\":\"\",\"inReviewEnabled\":false,\"inReviewRevisionsEnabled\":true}}],\"origin\":\"\",\"ownerIdentity\":\"b737a566-2e5e-46db-807a-df6ebddb134e\",\"owner\":[],\"postedDate\":\"February 18th, 2026\",\"published\":true,\"recentEditorialEvents\":[],\"rejectedJournal\":[],\"revision\":\"\",\"amendment\":\"\",\"status\":\"posted\",\"subjectAreas\":[{\"id\":63076762,\"name\":\"Infectious Diseases\"}],\"tags\":[],\"updatedAt\":\"2026-02-18T07:27:53+00:00\",\"versionOfRecord\":[],\"versionCreatedAt\":\"2026-02-18 07:27:53\",\"video\":\"\",\"vorDoi\":\"\",\"vorDoiUrl\":\"\",\"workflowStages\":[]},\"version\":\"v1\",\"identity\":\"rs-8901882\",\"journalConfig\":\"researchsquare\"},\"__N_SSP\":true},\"page\":\"/article/[identity]/[[...version]]\",\"query\":{\"redirect\":\"/article/rs-8901882\",\"identity\":\"rs-8901882\",\"version\":[\"v1\"]},\"buildId\":\"XKTyCvWXoU3ODBz1xrDgd\",\"isFallback\":false,\"isExperimentalCompile\":false,\"dynamicIds\":[84888],\"gssp\":true,\"scriptLoader\":[]}","source_license":"CC-BY-4.0","license_restricted":false}