Integrating Large Language Models Into UAE Community Pharmacies: Pharmacists′ Perspectives on Benefits, Concerns, and Implementation Barriers

High-quality discharge summaries are associated with improved patient outcomes, but contribute to clinical documentation burden. Large language models (LLMs) provide an opportunity to support physicians by drafting discharge summary narratives. To determine whether LLM-generated discharge summary narratives are of comparable quality and safety to those of physicians. This cross-sectional study conducted at the University of California, San Francisco included 100 randomly selected inpatient hospital medicine encounters of 3 to 6 days' duration between 2019 and 2022. The analysis took place in July 2024. A blinded evaluation of physician- and LLM-generated narratives was performed in duplicate by 22 attending physician reviewers. Narratives were reviewed for overall quality, reviewer preference, comprehensiveness, concision, coherence, and 3 error types (inaccuracies, omissions, and hallucinations). Each error individually, and each narrative overall, were assigned potential harmfulness scores ranging from 0 to 7 on an adapted Agency for Healthcare Research and Quality scale. Across 100 encounters, LLM- and physician-generated narratives were comparable in overall quality on a Likert scale ranging from 1 to 5 (higher scores indicate higher quality; mean [SD] score, 3.67 [0.49] vs 3.77 [0.57]; P = .21) and reviewer preference (χ2 = 5.2; P = .27). LLM-generated narratives were more concise (mean [SD] score, 4.01 [0.37] vs 3.70 [0.59]; P < .001) and more coherent (mean [SD] score, 4.16 [0.39] vs 4.01 [0.53]; P = .02) than their physician-generated counterparts, but less comprehensive (mean [SD] score, 3.72 [0.58] vs 4.13 [0.58]; P < .001). LLM-generated narratives contained more unique errors (mean [SD] errors per summary, 2.91 [2.54]) than physician-generated narratives (mean [SD] errors per summary, 1.82 [1.94]). There was no significant difference in the potential for harm between LLM- and physician-generated narratives across individual errors (mean [SD] of 1.35 [1.07] vs 1.34 [1.05]; P = .99), with 6 and 5 individual errors, respectively, with scores of 4 (potential for permanent harm) or greater. Both LLM- and physician-generated narratives had low overall potential for harm (scores <1 on a scale ranging from 0-7), with LLM-generated narratives scoring higher than physician narratives (mean [SD] score of 0.84 [0.98] vs 0.36 [0.70]; P < .001) and only 1 LLM-generated narrative (compared with 0 physician-generated narratives) scoring 4 or greater. In this cross-sectional study of 100 inpatient hospital medicine encounters, LLM-generated discharge summary narratives were of comparable quality, and were preferred equally, to those generated by physicians. LLM-generated narratives were more likely to contain errors but had low overall harmfulness scores. These results suggest that, in clinical practice, using such narratives after human review may provide a viable option for hospitalists.

This editorial explores the evolving and transformative role of large language models (LLMs) in enhancing the capabilities of virtual assistants (VAs) in the health care domain, highlighting recent research on the performance of VAs and LLMs in health care information sharing. Focusing on recent research, this editorial unveils the marked improvement in the accuracy and clinical relevance of responses from LLMs, such as GPT-4, compared to current VAs, especially in addressing complex health care inquiries, like those related to postpartum depression. The improved accuracy and clinical relevance with LLMs mark a paradigm shift in digital health tools and VAs. Furthermore, such LLM applications have the potential to dynamically adapt and be integrated into existing VA platforms, offering cost-effective, scalable, and inclusive solutions. These suggest a significant increase in the applicable range of VA applications, as well as the increased value, risk, and impact in health care, moving toward more personalized digital health ecosystems. However, alongside these advancements, it is necessary to develop and adhere to ethical guidelines, regulatory frameworks, governance principles, and privacy and safety measures. We need a robust interdisciplinary collaboration to navigate the complexities of safely and effectively integrating LLMs into health care applications, ensuring that these emerging technologies align with the diverse needs and ethical considerations of the health care domain.

<sec> <title>BACKGROUND</title> Substance use disorder (SUD) affects more than 48 million people in the US, but access to care is often constrained by factors such as stigma, financial barriers, and limited service availability. While digital health interventions have begun to mitigate these challenges, large language models (LLMs) such as GPT introduce new possibilities for scalable SUD recovery support. However, empirical evidence regarding LLM use in SUD contexts re-mains fragmented and unevenly evaluated. A scoping review is needed to synthesize cur-rent approaches, surface methodological and safety gaps, and inform future research. </sec> <sec> <title>OBJECTIVE</title> This scoping review synthesizes current evidence on the use of LLMs in SUD and addiction contexts. It examines how LLM-based tools are being applied across the full spectrum of care, identifies common tasks and data sources, evaluates risks and safety concerns, and maps early evidence of effectiveness and gaps that inform future research. </sec> <sec> <title>METHODS</title> We conducted a scoping review in accordance with PRISMA-ScR guidelines. We searched medical and technical databases for studies that reported empirical findings on LLMs in drug-related SUD contexts. Of 675 records identified, 492 were screened at the title and abstract level, 60 underwent full-text review, and 12 studies met inclusion criteria. </sec> <sec> <title>RESULTS</title> All included studies were published in 2024 or 2025, indicating a very recent and rapidly evolving evidence base. Most studies used public social media or online forums (8 of 12), while 3 studies used clinical data such as EHR, and 1 study used multimodal data that combined sensors, images, and chat logs. GPT-family models were the most frequently used (10 of 12), often evaluated alongside Llama or Mistral models. Across studies, LLM applications clustered into four task categories: text classification and extraction, response generation, conversational decision support, and predictive modeling using LLM-derived features. Classification and extraction tasks were most common and often showed im-proved performance compared with traditional machine learning or rule-based baselines. While general-purpose LLMs were prone to generating stigmatizing, biased, or clinically inappropriate content in SUD-related scenarios, it was also reported that targeted prompt engineering substantially reduced stigmatizing language and improved perceived ade-quacy and appropriateness of responses. However, safety evaluations remained limited in scope and depth. </sec> <sec> <title>CONCLUSIONS</title> LLM-based tools show emerging potential to support SUD-related research and care, particularly scalable classification and information extraction, conversational screening, and feature derivation from unstructured text or speech. Yet current work is predominantly preclinical, relying on public datasets and simulated prompts, with limited evidence on safety, clinical effectiveness, or real-world implementation in recovery settings. Responsible use in SUD contexts will require a safety-first development approach that integrates human-in-the-loop validation, systematic prompt engineering, and design constraints such as curated response libraries. It will also require efforts to strengthen LLM literacy among clinicians, peer specialists, and people in recovery. </sec>

In recent years, the landscape of digital biomedicine and healthcare has been reshaped due to the disruptive breakthroughs in AI-facilitated by tremendous data and high-performance computers, large language models (LLMs) have transformed information technology from accessing data to performing analytical tasks. While demonstrating unprecedented capabilities, LLMs have been found unreliable in tasks requiring factual knowledge and rigorous reasoning. Biomedicine and healthcare, as an important vertical domain rapidly benefitting from progress in AI, necessitates strict requirements on the accuracy, controllability, and interpretability of analytical models, posing critical challenges for LLMs. Despite recent studies addressing the hallucination problem of LLMs, research on empowering LLMs with the ability to plan, reason, and ground with explicit knowledge has also started to prosper, especially in the biomedicine and healthcare domain. On the other hand, biomedical data are enormous and notoriously complex, coming from various sources (e.g., biomedical knowledge bases, online literature, and hospitals) and bearing various modalities (e.g., tables, texts, images and time-series). Healthcare professionals have spent decades collecting, cleaning, and curating various types of data. The processes are extremely costly, producing various datasets with different data schemas, coding systems, and quality standards, many privately owned by the creators, making their integrative analysis and utilization through unified AI techniques still rather challenging. The generalizability of LLMs across different types of data endow them strong promises in automating the processing of large-scale complex healthcare data such as into unified knowledge graphs (KGs). Our goal in this survey is to systematically investigate and summarize recent studies on the unification of LLMs and KGs, towards fully utilizing the value of complex data, unleashing the power of generative AI, and expediting next-generation AI for biomedicine and healthcare applications.

Reinforcement Learning with Human Feedback (RLHF) is a methodology designed to align Large Language Models (LLMs) with human preferences, playing an important role in LLMs alignment. Despite its advantages, RLHF relies on human annotators to rank the text, which can introduce potential security vulnerabilities if any adversarial annotator (i.e., attackers) manipulates the ranking score by upranking any malicious text to steer the LLM adversarially. To assess the red-teaming of RLHF against human preference data poisoning, we propose RankPoison, a poisoning attack method on candidates' selection of preference rank flipping to reach certain malicious behaviors (e.g., generating longer sequences, which can increase the computational cost). With poisoned dataset generated by RankPoison, we can perform poisoning attacks on LLMs to generate longer tokens without hurting the original safety alignment performance. Moreover, applying RankPoison, we also successfully implement a backdoor attack where LLMs can generate longer answers under questions with the trigger word. Our findings highlight critical security challenges in RLHF, underscoring the necessity for more robust alignment methods for LLMs.

The digital health field's pursuit of massive, personalized healthcare continuously faces constraints from doctors' resources and capacity limitations. Recently, the emergence of large language models (LLMs), with their remarkable comprehension and processing abilities, has revolutionized digital health and enhanced massive, personalized healthcare. Although these LLMs have achieved significant advancements, they have also introduced inevitable hallucinations, which impact patient safety when used in massive applications. To address these challenges, this study proposes a digital hospital for a massive, personalized, reliable healthcare service named the Chat Chain-Brain-based Doctor (CHATCBD). In addition, this study transforms the LLM-based diagnostic process into a digital hospital architecture, designs a controllable AI agents framework, and develops a self-audit mechanism to enhance their reliability. The proposed CHATCBD uses blockchain technology to decentralize external regulation of the LLMs' personalized diagnoses. It introduces a blockchain-based personalized routing management mechanism to improve patient-centered decision-making and designs a blockchain-based audit framework based on a proposed mathematical model that ensures both the professionalism and honesty of audits, serving as a safety net for addressing LLM hallucinations. The results of extensive experiments conducted on 13 datasets from multiple perspectives demonstrate that the proposed CHATCBD system can significantly enhance the capabilities of LLMs in personalized healthcare.

Large language model (LLM) failures are commonly treated as undesirable artifacts arising from hallucination, misalignment, or insufficient training. In this work, we demonstrate that failure can instead be deliberately constructed, stabilized, and studied as a geometric property of constrained semantic spaces. We present a controlled experimental system in which an instruction-tuned local LLM is placed within a semantic manifold containing strong topic attractors, mandatory escalation rules, and physically enforced termination boundaries. Under these conditions, the model exhibits predictable attractor-driven spiraling behavior characterized by increasing epistemic certainty, decreasing evidentiary grounding, and inevitable semantic collapse. We argue that this behavior is not pathological but geometrically necessary given the imposed constraints. The resulting system functions as a runnable instantiation of Failure Geometry, providing both a diagnostic instrument for LLM behavior and a constructive framework for analyzing obsession, recursion, and collapse in generative systems.

Generative AI, powered by large language models, is transforming consumer health by enhancing health literacy and delivering personalized health education. However, ensuring clinical safety and effectiveness requires a robust digital health framework to address risks like misinformation and inequitable communication. This mini review examines current use cases for generative AI in consumer health education, highlights persistent challenges, and proposes a clinician-informed framework to evaluate safety, usability, and effectiveness. The RECAP model—Relevance, Evidence-based, Clarity, Adaptability, and Precision—offers a pragmatic lens to guide responsible implementation of AI in patient-facing tools. By connecting insights from past digital health innovations to the opportunities and pitfalls of large language models, this paper provides both context and direction for future development.

Artificial intelligence (AI) technologies, including machine learning (ML), deep learning, and large language models, are increasingly applied in medical diagnostics. In rhinology, these tools are being evaluated for tasks such as image interpretation, cytology classification, and clinical decision support. To systematically evaluate the application and diagnostic accuracy of AI technologies in rhinology, with a focus on clinical utility and implementation barriers. This review followed Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA) 2020 guidelines. Seventeen full-text studies were screened based on predefined eligibility criteria, focusing on AI applications with diagnostic metrics in rhinology. Data on AI type, diagnostic task, performance outcomes, and study quality were extracted and synthesized narratively. Twelve studies met the inclusion criteria. Image-based diagnostic tools using convolutional neural networks demonstrated high accuracy (81%-99%) in nasal polyp detection, cytology classification, and computed tomography (CT) scan interpretation. ML models using patient-reported data achieved accuracies of 74.5%-85.5% for chronic rhinosinusitis prediction. Large language models like ChatGPT and Gemini were evaluated for clinical question answering, with performance exceeding 80% in some domains. Risk of bias was moderate in most primary studies, and none reported clinical integration beyond prototype stages. AI exhibits promising diagnostic accuracy across several applications in rhinology. However, significant challenges persist, including limited validation, methodological heterogeneity, and lack of clinical implementation. Future research should focus on prospective trials, explainability, and regulatory frameworks to ensure safe integration into clinical workflows.

AI-assisted programming has gained significant traction in recent years, primarily driven by advances in large language model (LLM) technologies. This growth has resulted in the development of tools such as ChatGPT and GitHub Copilot. However, the use of AI models for code generation presents a critical challenge: The generated code is prone to security vulnerabilities, raising concerns within the domain of secure software development. This study examines the current research landscape on the security implications of LLM-generated code from a software engineering perspective. To this end, we conducted a multivocal literature review (MLR) in accordance with the guidelines outlined by Kitchenham et al. and Garousi et al. The search encompassed five academic sources: IEEE Xplore, ACM, ScienceDirect, Springer Link, and Wiley Online Library; as well as two grey literature sources: Google Scholar and Google. We supplemented the selection by means of backward and forward snowballing techniques. Through Quasi-Gold’s automated search, we identified 2796 peer-reviewed studies, from which we selected 48 primary studies. The analysis revealed seven distinct categories of security vulnerabilities in LLM-generated code, six mitigation strategies and best practices, and four tools recommended for improving security in AI-assisted coding. While research on this issue remains in its early stages, initial findings highlight the need for caution when integrating AI-powered code generation into software development, as it may introduce vulnerabilities at a higher rate than human-written code. The vulnerabilities, mitigation strategies, and tools identified in this study can serve as valuable resources for developers seeking to use LLM-based programming assistants more securely and responsibly.

Natural Language Processing (NLP) has seen a surge in recent years, especially with the introduction of transformer architectures, relying on the now famous self-attention mechanism. Especially, with the rise of Large Language Models (LLM), propelled by the appearance of ChatGPT in 2022, a new hope of extracting relevant information from text has emerged. In the meantime, natural language data have not often been used in risk, resilience, and reliability tasks. However, text data containing reliability-related information, that can be used to monitor health information regarding complex systems, are available in several and diverse shapes. Indeed, text data can either contain theoretical expert knowledge (technical reports, documentation, Failure Modes and Effects Analysis (FMEA)), or in-practice expert knowledge (incident reports, maintenance work orders), or in-practice non-expert knowledge (customer feedback, news articles). Critical infrastructures, such as nuclear powerplants, railway networks, or electrical power grids, are complex systems for which any failure would induce severe consequences affecting many people. Such systems have the advantage of serving many users, thus having many possible text sources from which technical information and past incident data can be mined for anticipating future failures and generating responses to catastrophic scenarios. The goal of this work is to develop methods and apply state-of-the-art NLP techniques to text data relating to critical infrastructures and failures, to (1) mine information from unstructured language data, and (2) structure the extracted information. Preliminary experiments were conducted on customer review data and incident reports, and show promising performance for failure detection from text data with transformers, as well as incident-related information extraction using LLMs.

In the field of confidentiality management, secret point recognition serves as an accurate and highly efficient approach. Due to the strict confidentiality requirements imposed on training data and the limited scope of data access, secret point recognition based on machine learning faces challenges in achieving good recognition results because of insufficient training data. This paper presents a secret point recognition algorithm using test-time augmentation (TTA) based on large language models. During training, a small amount of data was used to finetune the pretrained model. During inference, a large language model is used to paraphrase the input text, generating multiple expressions with different styles but the same meaning. These expressions are used as augmented data and are classified using a fine-tuned pre-trained classification model. The classification results were aggregated through an aggregation strategy to obtain the final prediction. To verify the effectiveness of the proposed method, we conducted experiments using simulated data. In the experiments, the ERNIE model was fine-tuned with a small amount of data. On this basis, TTA based on large language model was applied, and compared it with the BiLSTM and TextCNN methods. The experimental results demonstrate that this method can effectively improve the accuracy by 4.2 percent.

Introduction/Background: In-hospital cardiac arrest (IHCA) is experienced by approximately 300,000 patients annually in the US. While individual care teams can readily identify IHCA at the bedside, the reporting of these events after the event often lacks consistency and poses a major challenge for hospitals across the US. Accurate and timely reporting of IHCA events is crucial for facilitating QI initiatives, such as CPR quality review, optimizing team performance, and benchmarking IHCA outcomes. Hypothesis: Large language models (LLMs) can be used to identify IHCA events at the patient encounter-level by analyzing targeted note types from the inpatient record with high performance. Methods: Adult (≥18 years old) inpatient encounters at the Hospital of the University of Pennsylvania from 06/2018 to 03/2022 with a reported clinical emergency were identified from a QI database. Discharge summaries and notes associated with the encounters of interest were extracted from the EHR and deidentified using a natural language processing tool. Notes were reviewed by research staff to ascertain true IHCA events. Positive IHCA was defined as the loss of pulses followed by the delivery of CPR. An LLM, GPT4 v. 32K-Chat, was employed on the Penn Medicine Microsoft Azure Databricks environment to compare zero- and few-shot prompting methods to identify IHCA. Model performance was measured using accuracy, precision, recall and F1 scores. Results: Manual chart review determined 96.8% and 3.2% of cases to be IHCA+ and IHCA-, respectively. A balanced pilot dataset of positive and negative cases was created and the accuracy values across strategically sampled note types was checked. Using zero-shot learning with Chain-of-Thought (COT) prompting, the LLM performed with an accuracy of 86%, precision of 77%, recall of 100%, and F1-score of 87%. Using three-shot learning with COT, the LLM performed with an accuracy of 90%, precision of 83%, recall of 100%, and F1-score of 91%. Conclusions: This study demonstrates the potential efficacy of leveraging LLMs to automatically classify IHCA from strategically sampled note types using both zero and few-shot learning approaches. In the future, we will implement temporal sampling of note types and refine model parameters such as output token size and temperature. These findings suggest that such technologies can be effective in real-world clinical settings, providing a scalable solution to improve patient outcomes and quality improvement efforts.

e13657 Background: With demanding schedules, oncologists often struggle to stay updated on the latest clinical trial data from publications, especially without access to pre-curated databases. Traditionally, building annotated clinical trial databases required extensive time and manual effort. To address these challenges, we explored assembling an L-SLR library of clinical trial data using extractions generated by an agentic LLM system and evaluated the system’s accuracy and associated time savings. Methods: Agentic LLM systems are autonomous systems where multiple LLMs maintain control over how they accomplish tasks with no human input or supervised training. Our system used three OpenAI LLM models (o1, o3, o1-mini) in a matrix of processes to emulate trained human experts by following an annotation manual, subdividing complex processes into smaller subtasks, and documenting its reasoning for traceable results. This methodology is inspired by the recent development from DeepSeek reinforcement learning and reasoning capabilities. Annotations were created for 4 review variables (population, intervention/comparator, outcome, study design) and 32 extraction variables, including clinical TNM staging, histology, biomarker, associated risk factors, treatment path (line, prior therapy etc.), interventions, intervention type; study randomization, phase and sample size; analysis type, follow-up period; reported outcomes (median and landmark overall survival (OS), progression-free survival (PFS) and other progression measures, response data, quality of life data); subgroup analyses, safety/toxicity. Accuracy of review and extraction was evaluated on publications in three cancer types: NSCLC, PC, BC compared to human results. Results: Our agentic LLM system generated annotations for 4 review variables for 19,407 publications (6,916 NSCLC, 6,978 PC, 5,513 BC) publications, and 32 extraction variables for 2,424 (1,356 NSCLC, 587 PC, 481 BC) publications. Accuracy for the review variables ranged from 93.8% to 97.2%. For extraction variables, accuracy exceeded 90% for all variables (91.5%-99.1%) with 50% of variables above 95%. Our system completed the annotations in 5.34 hours, compared to an estimated 727.39 hours by trained human researchers, resulting in 99.27% time savings. Conclusions: Our living SLR system can accurately review and extract clinical trial publications with performance comparable to human experts. This level of accuracy highlights our system’s potential to deliver real-time clinical data, empowering oncologists to make more informed treatment decisions, with the hopes of ultimately improving patient outcomes.

Large Language Models (LLMs) have surged in popularity due to their impressive ability to generate human-like text. Despite the widespread use of large language models, there is a growing concern about their disregard for human ethics and their potential to produce harmful content. While many LLMs are aligned with safeguards, there is a category of prompt injection attacks known as jailbreaks, specifically designed to bypass these protections and generate malicious output. Despite extensive research on novel jailbreak attacks and potential defenses, there is limited exploration into accurately evaluating the success of these attacks. In this paper, we introduce seven evaluation methods used in re- search to determine the effectiveness of jailbreak attempts. We conduct a comprehensive analysis of these seven methods with a particular focus on their accuracy. Our research aims to advance the discussion on improving the safety and alignment of LLMs with human values and to contribute to the development of more robust and secure LLM-based applications. Code is available at github.com/cenacle e18-4yp-An-Empirical-Study- On-Prompt-Injection-Attacks-And-Defenses. Due to the weaknesses of these basic evaluation methods, there is a risk of misrepresenting the actual effectiveness of jailbreak attacks and the security vulnerabilities of the models. Therefore, this research provides a comprehensive analysis of the often-overlooked limitations of each evaluation method and their accuracy. Our goal is to lay the foundation for the development of more standardized, reliable, and measurable evaluation metrics to determine the success of an attack. This will lay the foundation for future security research, while enabling the creation of more secure and user-friendly LLM applications.