Improving Prediction of Complications Post-Proton Therapy in Lung Cancer Using Large Language Models and Meta-Analysis.

PURPOSE: This study enhances the efficiency of predicting complications in lung cancer patients receiving proton therapy by utilizing large language models (LLMs) and meta-analytical techniques for literature quality assessment. MATERIALS AND METHODS: We integrated systematic reviews with LLM evaluations, sourcing studies from Web of Science, PubMed, and Scopus, managed via EndNote X20. Inclusion and exclusion criteria ensured literature relevance. Techniques included meta-analysis, heterogeneity assessment using Cochran's Q test and I2 statistics, and subgroup analyses for different complications. Quality and bias risk were assessed using the PROBAST tool and further analyzed with models such as ChatGPT-4, Llama2-13b, and Llama3-8b. Evaluation metrics included AUC, accuracy, precision, recall, F1 score, and time efficiency (WPM). RESULTS: The meta-analysis revealed an overall effect size of 0.78 for model predictions, with high heterogeneity observed (I2 = 72.88%, P < 0.001). Subgroup analysis for radiation-induced esophagitis and pneumonitis revealed predictive effect sizes of 0.79 and 0.77, respectively, with a heterogeneity index (I2) of 0%, indicating that there were no significant differences among the models in predicting these specific complications. A literature assessment using LLMs demonstrated that ChatGPT-4 achieved the highest accuracy at 90%, significantly outperforming the Llama3 and Llama2 models, which had accuracies ranging from 44% to 62%. Additionally, LLM evaluations were conducted 3229 times faster than manual assessments were, markedly enhancing both efficiency and accuracy. The risk assessment results identified nine studies as high risk, three as low risk, and one as unknown, confirming the robustness of the ChatGPT-4 across various evaluation metrics. CONCLUSION: This study demonstrated that the integration of large language models with meta-analysis techniques can significantly increase the efficiency of literature evaluations and reduce the time required for assessments, confirming that there are no significant differences among models in predicting post proton therapy complications in lung cancer patients.

Using Advanced AI to Improve Predictions of Treatment Side Effects in Lung Cancer: This research uses cutting-edge artificial intelligence (AI) techniques, including large language models like ChatGPT-4, to better predict potential side effects in lung cancer patients undergoing proton therapy. By analyzing extensive scientific literature quickly and accurately, this approach has proven to enhance the evaluation process, making it faster and more reliable in foreseeing complications from treatments.

Large language model triaging of simulated nephrology patient inbox messages.

Background: Efficient triage of patient communications is crucial for timely medical attention and improved care. This study evaluates ChatGPT's accuracy in categorizing nephrology patient inbox messages, assessing its potential in outpatient settings. Methods: One hundred and fifty simulated patient inbox messages were created based on cases typically encountered in everyday practice at a nephrology outpatient clinic. These messages were triaged as non-urgent, urgent, and emergent by two nephrologists. The messages were then submitted to ChatGPT-4 for independent triage into the same categories. The inquiry process was performed twice with a two-week period in between. ChatGPT responses were graded as correct (agreement with physicians), overestimation (higher priority), or underestimation (lower priority). Results: In the first trial, ChatGPT correctly triaged 140 (93%) messages, overestimated the priority of 4 messages (3%), and underestimated the priority of 6 messages (4%). In the second trial, it correctly triaged 140 (93%) messages, overestimated the priority of 9 (6%), and underestimated the priority of 1 (1%). The accuracy did not depend on the urgency level of the message (p = 0.19). The internal agreement of ChatGPT responses was 92% with an intra-rater Kappa score of 0.88. Conclusion: ChatGPT-4 demonstrated high accuracy in triaging nephrology patient messages, highlighting the potential for AI-driven triage systems to enhance operational efficiency and improve patient care in outpatient clinics.

Unveiling the risks of ChatGPT in diagnostic surgical pathology.

ChatGPT, an AI capable of processing and generating human-like language, has been studied in medical education and care, yet its potential in histopathological diagnosis remains unexplored. This study evaluates ChatGPT's reliability in addressing pathology-related diagnostic questions across ten subspecialties and its ability to provide scientific references. We crafted five clinico-pathological scenarios per subspecialty, simulating a pathologist using ChatGPT to refine differential diagnoses. Each scenario, aligned with current diagnostic guidelines and validated by expert pathologists, was posed as open-ended or multiple-choice questions, either requesting scientific references or not. Outputs were assessed by six pathologists according to. (1) usefulness in supporting the diagnosis and (2) absolute number of errors. We used directed acyclic graphs and structural causal models to determine the effect of each scenario type, field, question modality, and pathologist evaluation. We yielded 894 evaluations. ChatGPT provided useful answers in 62.2% of cases, and 32.1% of outputs contained no errors, while the remaining had at least one error. ChatGPT provided 214 bibliographic references: 70.1% correct, 12.1% inaccurate, and 17.8% non-existing. Scenario variability had the greatest impact on ratings, and latent knowledge across fields showed minimal variation. Although ChatGPT provided useful responses in one-third of cases, the frequency of errors and variability underscores its inadequacy for routine diagnostic use and highlights the need for discretion as a support tool. Imprecise referencing also suggests caution as a self-learning tool. It is essential to recognize the irreplaceable role of human experts in synthesizing images, clinical data, and experience for the intricate task of histopathological diagnosis.

Can GPT-3.5 generate and code discharge summaries?

OBJECTIVES: The aim of this study was to investigate GPT-3.5 in generating and coding medical documents with International Classification of Diseases (ICD)-10 codes for data augmentation on low-resource labels. MATERIALS AND METHODS: Employing GPT-3.5 we generated and coded 9606 discharge summaries based on lists of ICD-10 code descriptions of patients with infrequent (or generation) codes within the MIMIC-IV dataset. Combined with the baseline training set, this formed an augmented training set. Neural coding models were trained on baseline and augmented data and evaluated on an MIMIC-IV test set. We report micro- and macro-F1 scores on the full codeset, generation codes, and their families. Weak Hierarchical Confusion Matrices determined within-family and outside-of-family coding errors in the latter codesets. The coding performance of GPT-3.5 was evaluated on prompt-guided self-generated data and real MIMIC-IV data. Clinicians evaluated the clinical acceptability of the generated documents. RESULTS: Data augmentation results in slightly lower overall model performance but improves performance for the generation candidate codes and their families, including 1 absent from the baseline training data. Augmented models display lower out-of-family error rates. GPT-3.5 identifies ICD-10 codes by their prompted descriptions but underperforms on real data. Evaluators highlight the correctness of generated concepts while suffering in variety, supporting information, and narrative. DISCUSSION AND CONCLUSION: While GPT-3.5 alone given our prompt setting is unsuitable for ICD-10 coding, it supports data augmentation for training neural models. Augmentation positively affects generation code families but mainly benefits codes with existing examples. Augmentation reduces out-of-family errors. Documents generated by GPT-3.5 state prompted concepts correctly but lack variety, and authenticity in narratives.

Predicting Differentially Methylated Cytosines in TET and DNMT3 Knockout Mutants via a Large Language Model.

DNA cytosine methylation is an epigenetic marker which regulates many cellular processes. Mammalian genomes typically maintain consistent methylation patterns over time, except in specific regulatory regions like promoters and certain types of enhancers. The dynamics of DNA methylation is controlled by a complex cellular machinery, in which the enzymes DNMT3 and TET play a major role. This study explores the identification of differentially methylated cytosines (DMCs) in TET and DNMT3 knockout mutants in mice and human embryonic stem cells. We investigate (i) whether a large language model can be trained to recognize DMCs in human and mouse from the sequence surrounding the cytosine of interest, (ii) whether a classifier trained on human knockout data can predict DMCs in the mouse genome (and vice versa), (iii) whether a classifier trained on DNMT3 knockout can predict DMCs for TET knockout (and vice versa). Our study identifies statistically significant motifs associated with the prediction of DMCs each mutant, casting a new light on the understanding of DNA methylation dynamics in stem cells. Our software tool is available at https://github.com/ucrbioinfo/dmc_prediction.

Enhancing Medical Interview Skills Through AI-Simulated Patient Interactions: Nonrandomized Controlled Trial.

BACKGROUND: Medical interviewing is a critical skill in clinical practice, yet opportunities for practical training are limited in Japanese medical schools, necessitating urgent measures. Given advancements in artificial intelligence (AI) technology, its application in the medical field is expanding. However, reports on its application in medical interviews in medical education are scarce. OBJECTIVE: This study aimed to investigate whether medical students' interview skills could be improved by engaging with AI-simulated patients using large language models, including the provision of feedback. METHODS: This nonrandomized controlled trial was conducted with fourth-year medical students in Japan. A simulation program using large language models was provided to 35 students in the intervention group in 2023, while 110 students from 2022 who did not participate in the intervention were selected as the control group. The primary outcome was the score on the Pre-Clinical Clerkship Objective Structured Clinical Examination (pre-CC OSCE), a national standardized clinical skills examination, in medical interviewing. Secondary outcomes included surveys such as the Simulation-Based Training Quality Assurance Tool (SBT-QA10), administered at the start and end of the study. RESULTS: The AI intervention group showed significantly higher scores on medical interviews than the control group (AI group vs control group: mean 28.1, SD 1.6 vs 27.1, SD 2.2; P=.01). There was a trend of inverse correlation between the SBT-QA10 and pre-CC OSCE scores (regression coefficient -2.0 to -2.1). No significant safety concerns were observed. CONCLUSIONS: Education through medical interviews using AI-simulated patients has demonstrated safety and a certain level of educational effectiveness. However, at present, the educational effects of this platform on nonverbal communication skills are limited, suggesting that it should be used as a supplementary tool to traditional simulation education.

Impact of a Digital Scribe System on Clinical Documentation Time and Quality: Usability Study.

BACKGROUND: Physicians spend approximately half of their time on administrative tasks, which is one of the leading causes of physician burnout and decreased work satisfaction. The implementation of natural language processing-assisted clinical documentation tools may provide a solution. OBJECTIVE: This study investigates the impact of a commercially available Dutch digital scribe system on clinical documentation efficiency and quality. METHODS: Medical students with experience in clinical practice and documentation (n=22) created a total of 430 summaries of mock consultations and recorded the time they spent on this task. The consultations were summarized using 3 methods: manual summaries, fully automated summaries, and automated summaries with manual editing. We then randomly reassigned the summaries and evaluated their quality using a modified version of the Physician Documentation Quality Instrument (PDQI-9). We compared the differences between the 3 methods in descriptive statistics, quantitative text metrics (word count and lexical diversity), the PDQI-9, Recall-Oriented Understudy for Gisting Evaluation scores, and BERTScore. RESULTS: The median time for manual summarization was 202 seconds against 186 seconds for editing an automatic summary. Without editing, the automatic summaries attained a poorer PDQI-9 score than manual summaries (median PDQI-9 score 25 vs 31, P<.001, ANOVA test). Automatic summaries were found to have higher word counts but lower lexical diversity than manual summaries (P<.001, independent t test). The study revealed variable impacts on PDQI-9 scores and summarization time across individuals. Generally, students viewed the digital scribe system as a potentially useful tool, noting its ease of use and time-saving potential, though some criticized the summaries for their greater length and rigid structure. CONCLUSIONS: This study highlights the potential of digital scribes in improving clinical documentation processes by offering a first summary draft for physicians to edit, thereby reducing documentation time without compromising the quality of patient records. Furthermore, digital scribes may be more beneficial to some physicians than to others and could play a role in improving the reusability of clinical documentation. Future studies should focus on the impact and quality of such a system when used by physicians in clinical practice.

Extracting lung cancer staging descriptors from pathology reports: A generative language model approach.

BACKGROUND: In oncology, electronic health records contain textual key information for the diagnosis, staging, and treatment planning of patients with cancer. However, text data processing requires a lot of time and effort, which limits the utilization of these data. Recent advances in natural language processing (NLP) technology, including large language models, can be applied to cancer research. Particularly, extracting the information required for the pathological stage from surgical pathology reports can be utilized to update cancer staging according to the latest cancer staging guidelines. OBJECTIVES: This study has two main objectives. The first objective is to evaluate the performance of extracting information from text-based surgical pathology reports and determining pathological stages based on the extracted information using fine-tuned generative language models (GLMs) for patients with lung cancer. The second objective is to determine the feasibility of utilizing relatively small GLMs for information extraction in a resource-constrained computing environment. METHODS: Lung cancer surgical pathology reports were collected from the Common Data Model database of Seoul National University Bundang Hospital (SNUBH), a tertiary hospital in Korea. We selected 42 descriptors necessary for tumor-node (TN) classification based on these reports and created a gold standard with validation by two clinical experts. The pathology reports and gold standard were used to generate prompt-response pairs for training and evaluating GLMs which then were used to extract information required for staging from pathology reports. RESULTS: We evaluated the information extraction performance of six trained models as well as their performance in TN classification using the extracted information. The Deductive Mistral-7B model, which was pre-trained with the deductive dataset, showed the best performance overall, with an exact match ratio of 92.24% in the information extraction problem and an accuracy of 0.9876 (predicting T and N classification concurrently) in classification. CONCLUSION: This study demonstrated that training GLMs with deductive datasets can improve information extraction performance, and GLMs with a relatively small number of parameters at approximately seven billion can achieve high performance in this problem. The proposed GLM-based information extraction method is expected to be useful in clinical decision-making support, lung cancer staging and research.

Can large language models be new supportive tools in coronary computed tomography angiography reporting?

The advent of large language models (LLMs) marks a transformative leap in natural language processing, offering unprecedented potential in radiology, particularly in enhancing the accuracy and efficiency of coronary artery disease (CAD) diagnosis. While previous studies have explored the capabilities of specific LLMs like ChatGPT in cardiac imaging, a comprehensive evaluation comparing multiple LLMs in the context of CAD-RADS 2.0 has been lacking. This study addresses this gap by assessing the performance of various LLMs, including ChatGPT 4, ChatGPT 4o, Claude 3 Opus, Gemini 1.5 Pro, Mistral Large, Meta Llama 3 70B, and Perplexity Pro, in answering 30 multiple-choice questions derived from the CAD-RADS 2.0 guidelines. Our findings reveal that ChatGPT 4o achieved the highest accuracy at 100 %, with ChatGPT 4 and Claude 3 Opus closely following at 96.6 %. Other models, including Mistral Large, Perplexity Pro, Meta Llama 3 70B, and Gemini 1.5 Pro, also demonstrated commendable performance, though with slightly lower accuracy ranging from 90 % to 93.3 %. This study underscores the proficiency of current LLMs in understanding and applying CAD-RADS 2.0, suggesting their potential to significantly enhance radiological reporting and patient care in coronary artery disease. The variations in model performance highlight the need for further research, particularly in evaluating the visual diagnostic capabilities of LLMs-a critical component of radiology practice. This study provides a foundational comparison of LLMs in CAD-RADS 2.0 and sets the stage for future investigations into their broader applications in radiology, emphasizing the importance of integrating both text-based and visual knowledge for optimal clinical outcomes.

Perceptions and expectations of an artificially intelligent physical activity digital assistant - A focus group study.

Artificially intelligent physical activity digital assistants that use the full spectrum of machine learning capabilities have not yet been developed and examined. This study aimed to explore potential users' perceptions and expectations of using such a digital assistant. Six 90-min online focus group meetings (n = 45 adults) were conducted. Meetings were recorded, transcribed and thematically analysed. Participants embraced the idea of a 'digital assistant' providing physical activity support. Participants indicated they would like to receive notifications from the digital assistant, but did not agree on the number, timing, tone and content of notifications. Likewise, they indicated that the digital assistant's personality and appearance should be customisable. Participants understood the need to provide information to the digital assistant to allow for personalisation, but varied greatly in the extent of information that they were willing to provide. Privacy issues aside, participants embraced the idea of using artificial intelligence or machine learning in return for a more functional and personal digital assistant. In sum, participants were ready for an artificially intelligent physical activity digital assistant but emphasised a need to personalise or customise nearly every feature of the application. This poses challenges in terms of cost and complexity of developing the application.

Driving assistant using generative AI pre-generated messages in simulator-based driving assessment: A step towards low-cost simulator-based driving assessment.

This paper presents a novel approach for a low-cost simulator-based driving assessment system incorporating a speech-based assistant, using pre-generated messages from Generative AI to achieve real-time interaction during the assessment. Simulator-based assessment is a crucial apparatus in the research toolkit for various fields. Traditional assessment approaches, like on-road evaluation, though reliable, can be risky, costly, and inaccessible. Simulator-based assessment using stationary driving simulators offers a safer evaluation and can be tailored to specific needs. However, these simulators are often only available to research-focused institutions due to their cost. To address this issue, our study proposes a system with the aforementioned properties aiming to enhance drivers' situational awareness, and foster positive emotional states, i.e., high valence and medium arousal, while assessing participants to prevent subpar performers from proceeding to the next stages of assessment and/or rehabilitation. In addition, this study introduces the speech-based assistant which provides timely guidance adaptable to the ever-changing context of the driving environment and vehicle state. The study's preliminary outcomes reveal encouraging progress, highlighting improved driving performance and positive emotional states when participants are engaged with the assistant during the assessment.

ChatGPT Use Among Pediatric Health Care Providers: Cross-Sectional Survey Study.

BACKGROUND: The public launch of OpenAI's ChatGPT platform generated immediate interest in the use of large language models (LLMs). Health care institutions are now grappling with establishing policies and guidelines for the use of these technologies, yet little is known about how health care providers view LLMs in medical settings. Moreover, there are no studies assessing how pediatric providers are adopting these readily accessible tools. OBJECTIVE: The aim of this study was to determine how pediatric providers are currently using LLMs in their work as well as their interest in using a Health Insurance Portability and Accountability Act (HIPAA)-compliant version of ChatGPT in the future. METHODS: A survey instrument consisting of structured and unstructured questions was iteratively developed by a team of informaticians from various pediatric specialties. The survey was sent via Research Electronic Data Capture (REDCap) to all Boston Children's Hospital pediatric providers. Participation was voluntary and uncompensated, and all survey responses were anonymous. RESULTS: Surveys were completed by 390 pediatric providers. Approximately 50% (197/390) of respondents had used an LLM; of these, almost 75% (142/197) were already using an LLM for nonclinical work and 27% (52/195) for clinical work. Providers detailed the various ways they are currently using an LLM in their clinical and nonclinical work. Only 29% (n=105) of 362 respondents indicated that ChatGPT should be used for patient care in its present state; however, 73.8% (273/368) reported they would use a HIPAA-compliant version of ChatGPT if one were available. Providers' proposed future uses of LLMs in health care are described. CONCLUSIONS: Despite significant concerns and barriers to LLM use in health care, pediatric providers are already using LLMs at work. This study will give policy makers needed information about how providers are using LLMs clinically.

Practical Applications of Large Language Models for Health Care Professionals and Scientists.

Unlabelled: With the popularization of large language models (LLMs), strategies for their effective and safe usage in health care and research have become increasingly pertinent. Despite the growing interest and eagerness among health care professionals and scientists to exploit the potential of LLMs, initial attempts may yield suboptimal results due to a lack of user experience, thus complicating the integration of artificial intelligence (AI) tools into workplace routine. Focusing on scientists and health care professionals with limited LLM experience, this viewpoint article highlights and discusses 6 easy-to-implement use cases of practical relevance. These encompass customizing translations, refining text and extracting information, generating comprehensive overviews and specialized insights, compiling ideas into cohesive narratives, crafting personalized educational materials, and facilitating intellectual sparring. Additionally, we discuss general prompting strategies and precautions for the implementation of AI tools in biomedicine. Despite various hurdles and challenges, the integration of LLMs into daily routines of physicians and researchers promises heightened workplace productivity and efficiency.

Prompt matters: evaluation of large language model chatbot responses related to Peyronie's disease.

Introduction: Despite direct access to clinicians through the electronic health record, patients are increasingly turning to the internet for information related to their health, especially with sensitive urologic conditions such as Peyronie's disease (PD). Large language model (LLM) chatbots are a form of artificial intelligence that rely on user prompts to mimic conversation, and they have shown remarkable capabilities. The conversational nature of these chatbots has the potential to answer patient questions related to PD; however, the accuracy, comprehensiveness, and readability of these LLMs related to PD remain unknown. Aims: To assess the quality and readability of information generated from 4 LLMs with searches related to PD; to see if users could improve responses; and to assess the accuracy, completeness, and readability of responses to artificial preoperative patient questions sent through the electronic health record prior to undergoing PD surgery. Methods: The National Institutes of Health's frequently asked questions related to PD were entered into 4 LLMs, unprompted and prompted. The responses were evaluated for overall quality by the previously validated DISCERN questionnaire. Accuracy and completeness of LLM responses to 11 presurgical patient messages were evaluated with previously accepted Likert scales. All evaluations were performed by 3 independent reviewers in October 2023, and all reviews were repeated in April 2024. Descriptive statistics and analysis were performed. Results: Without prompting, the quality of information was moderate across all LLMs but improved to high quality with prompting. LLMs were accurate and complete, with an average score of 5.5 of 6.0 (SD, 0.8) and 2.8 of 3.0 (SD, 0.4), respectively. The average Flesch-Kincaid reading level was grade 12.9 (SD, 2.1). Chatbots were unable to communicate at a grade 8 reading level when prompted, and their citations were appropriate only 42.5% of the time. Conclusion: LLMs may become a valuable tool for patient education for PD, but they currently rely on clinical context and appropriate prompting by humans to be useful. Unfortunately, their prerequisite reading level remains higher than that of the average patient, and their citations cannot be trusted. However, given their increasing uptake and accessibility, patients and physicians should be educated on how to interact with these LLMs to elicit the most appropriate responses. In the future, LLMs may reduce burnout by helping physicians respond to patient messages.

Extracting Critical Information from Unstructured Clinicians' Notes Data to Identify Dementia Severity Using a Rule-Based Approach: Feasibility Study.

BACKGROUND: The severity of Alzheimer disease and related dementias (ADRD) is rarely documented in structured data fields in electronic health records (EHRs). Although this information is important for clinical monitoring and decision-making, it is often undocumented or "hidden" in unstructured text fields and not readily available for clinicians to act upon. OBJECTIVE: We aimed to assess the feasibility and potential bias in using keywords and rule-based matching for obtaining information about the severity of ADRD from EHR data. METHODS: We used EHR data from a large academic health care system that included patients with a primary discharge diagnosis of ADRD based on ICD-9 (International Classification of Diseases, Ninth Revision) and ICD-10 (International Statistical Classification of Diseases, Tenth Revision) codes between 2014 and 2019. We first assessed the presence of ADRD severity information and then the severity of ADRD in the EHR. Clinicians' notes were used to determine the severity of ADRD based on two criteria: (1) scores from the Mini Mental State Examination and Montreal Cognitive Assessment and (2) explicit terms for ADRD severity (eg, "mild dementia" and "advanced Alzheimer disease"). We compiled a list of common ADRD symptoms, cognitive test names, and disease severity terms, refining it iteratively based on previous literature and clinical expertise. Subsequently, we used rule-based matching in Python using standard open-source data analysis libraries to identify the context in which specific words or phrases were mentioned. We estimated the prevalence of documented ADRD severity and assessed the performance of our rule-based algorithm. RESULTS: We included 9115 eligible patients with over 65,000 notes from the providers. Overall, 22.93% (2090/9115) of patients were documented with mild ADRD, 20.87% (1902/9115) were documented with moderate or severe ADRD, and 56.20% (5123/9115) did not have any documentation of the severity of their ADRD. For the task of determining the presence of any ADRD severity information, our algorithm achieved an accuracy of >95%, specificity of >95%, sensitivity of >90%, and an F1-score of >83%. For the specific task of identifying the actual severity of ADRD, the algorithm performed well with an accuracy of >91%, specificity of >80%, sensitivity of >88%, and F1-score of >92%. Comparing patients with mild ADRD to those with more advanced ADRD, the latter group tended to contain older, more likely female, and Black patients, and having received their diagnoses in primary care or in-hospital settings. Relative to patients with undocumented ADRD severity, those with documented ADRD severity had a similar distribution in terms of sex, race, and rural or urban residence. CONCLUSIONS: Our study demonstrates the feasibility of using a rule-based matching algorithm to identify ADRD severity from unstructured EHR report data. However, it is essential to acknowledge potential biases arising from differences in documentation practices across various health care systems.

Diagnostic Performance of GPT-4o and Claude 3 Opus in Determining Causes of Death From Medical Histories and Postmortem CT Findings.

INTRODUCTION: This study evaluates the diagnostic performance of the latest large language models (LLMs), GPT-4o (OpenAI, San Francisco, CA, USA) and Claude 3 Opus (Anthropic, San Francisco, CA, USA), in determining causes of death from medical histories and postmortem CT findings. METHODS: We included 100 adult cases whose postmortem CT scans were diagnosable for the causes of death using the gold standard of autopsy results. Their medical histories and postmortem CT findings were compiled, and clinical and imaging diagnoses of both the underlying and immediate causes of death, as well as their personal information, were carefully separated from the database to be shown to the LLMs. Both GPT-4o and Claude 3 Opus generated the top three differential diagnoses for each of the underlying or immediate causes of death based on the following three prompts: 1) medical history only; 2) postmortem CT findings only; and 3) both medical history and postmortem CT findings. The diagnostic performance of the LLMs was compared using McNemar's test. RESULTS: For the underlying cause of death, GPT-4o achieved primary diagnostic accuracy rates of 78%, 72%, and 78%, while Claude 3 Opus achieved 72%, 56%, and 75% for prompts 1, 2, and 3, respectively. Including any of the top three differential diagnoses, GPT-4o's accuracy rates were 92%, 90%, and 92%, while Claude 3 Opus's rates were 93%, 69%, and 93% for prompts 1, 2, and 3, respectively. For the immediate cause of death, GPT-4o's primary diagnostic accuracy rates were 55%, 58%, and 62%, while Claude 3 Opus's rates were 60%, 62%, and 63% for prompts 1,2, and 3, respectively. For any of the top three differential diagnoses, GPT-4o's accuracy rates were 88% for prompt 1 and 91% for prompts 2 and 3, whereas Claude 3 Opus's rates were 92% for all three prompts. Significant differences between the models were observed for prompt two in diagnosing the underlying cause of death (p = 0.03 and <0.01 for the primary and top three differential diagnoses, respectively). CONCLUSION: Both GPT-4o and Claude 3 Opus demonstrated relatively high performance in diagnosing both the underlying and immediate causes of death using medical histories and postmortem CT findings.

Assessing the utility of large language models for phenotype-driven gene prioritization in the diagnosis of rare genetic disease.

Phenotype-driven gene prioritization is fundamental to diagnosing rare genetic disorders. While traditional approaches rely on curated knowledge graphs with phenotype-gene relations, recent advancements in large language models (LLMs) promise a streamlined text-to-gene solution. In this study, we evaluated five LLMs, including two generative pre-trained transformers (GPT) series and three Llama2 series, assessing their performance across task completeness, gene prediction accuracy, and adherence to required output structures. We conducted experiments, exploring various combinations of models, prompts, phenotypic input types, and task difficulty levels. Our findings revealed that the best-performed LLM, GPT-4, achieved an average accuracy of 17.0% in identifying diagnosed genes within the top 50 predictions, which still falls behind traditional tools. However, accuracy increased with the model size. Consistent results were observed over time, as shown in the dataset curated after 2023. Advanced techniques such as retrieval-augmented generation (RAG) and few-shot learning did not improve the accuracy. Sophisticated prompts were more likely to enhance task completeness, especially in smaller models. Conversely, complicated prompts tended to decrease output structure compliance rate. LLMs also achieved better-than-random prediction accuracy with free-text input, though performance was slightly lower than with standardized concept input. Bias analysis showed that highly cited genes, such as BRCA1, TP53, and PTEN, are more likely to be predicted. Our study provides valuable insights into integrating LLMs with genomic analysis, contributing to the ongoing discussion on their utilization in clinical workflows.

Semantic search using protein large language models detects class II microcins in bacterial genomes.

Class II microcins are antimicrobial peptides that have shown some potential as novel antibiotics. However, to date, only 10 class II microcins have been described, and the discovery of novel microcins has been hampered by their short length and high sequence divergence. Here, we ask if we can use numerical embeddings generated by protein large language models to detect microcins in bacterial genome assemblies and whether this method can outperform sequence-based methods such as BLAST. We find that embeddings detect known class II microcins much more reliably than does BLAST and that any two microcins tend to have a small distance in embedding space even though they typically are highly diverged at the sequence level. In data sets of Escherichia coli, Klebsiella spp., and Enterobacter spp. genomes, we further find novel putative microcins that were previously missed by sequence-based search methods. IMPORTANCE: Antibiotic resistance is becoming an increasingly serious problem in modern medicine, but the development pipeline for conventional antibiotics is not promising. Therefore, alternative approaches to combat bacterial infections are urgently needed. One such approach may be to employ naturally occurring antibacterial peptides produced by bacteria to kill competing bacteria. A promising class of such peptides are class II microcins. However, only a small number of class II microcins have been discovered to date, and the discovery of further such microcins has been hampered by their high sequence divergence and short length, which can cause sequence-based search methods to fail. Here, we demonstrate that a more robust method for microcin discovery can be built on the basis of a protein large language model, and we use this method to identify several putative novel class II microcins.

Empowering Graph Neural Network-Based Computational Drug Repositioning with Large Language Model-Inferred Knowledge Representation.

Computational drug repositioning, through predicting drug-disease associations (DDA), offers significant potential for discovering new drug indications. Current methods incorporate graph neural networks (GNN) on drug-disease heterogeneous networks to predict DDAs, achieving notable performances compared to traditional machine learning and matrix factorization approaches. However, these methods depend heavily on network topology, hampered by incomplete and noisy network data, and overlook the wealth of biomedical knowledge available. Correspondingly, large language models (LLMs) excel in graph search and relational reasoning, which can possibly enhance the integration of comprehensive biomedical knowledge into drug and disease profiles. In this study, we first investigate the contribution of LLM-inferred knowledge representation in drug repositioning and DDA prediction. A zero-shot prompting template was designed for LLM to extract high-quality knowledge descriptions for drug and disease entities, followed by embedding generation from language models to transform the discrete text to continual numerical representation. Then, we proposed LLM-DDA with three different model architectures (LLM-DDANode Feat, LLM-DDADual GNN, LLM-DDAGNN-AE) to investigate the best fusion mode for LLM-based embeddings. Extensive experiments on four DDA benchmarks show that, LLM-DDAGNN-AE achieved the optimal performance compared to 11 baselines with the overall relative improvement in AUPR of 23.22%, F1-Score of 17.20%, and precision of 25.35%. Meanwhile, selected case studies of involving Prednisone and Allergic Rhinitis highlighted the model's capability to identify reliable DDAs and knowledge descriptions, supported by existing literature. This study showcases the utility of LLMs in drug repositioning with its generality and applicability in other biomedical relation prediction tasks.