Search

BACKGROUND: Ontologies are fundamental components of informatics infrastructure in domains such as biomedical, environmental, and food sciences, representing consensus knowledge in an accurate and computable form. However, their construction and maintenance demand substantial resources and necessitate substantial collaboration between domain experts, curators, and ontology experts. We present Dynamic Retrieval Augmented Generation of Ontologies using AI (DRAGON-AI), an ontology generation method employing Large Language Models (LLMs) and Retrieval Augmented Generation (RAG). DRAGON-AI can generate textual and logical ontology components, drawing from existing knowledge in multiple ontologies and unstructured text sources. RESULTS: We assessed performance of DRAGON-AI on de novo term construction across ten diverse ontologies, making use of extensive manual evaluation of results. Our method has high precision for relationship generation, but has slightly lower precision than from logic-based reasoning. Our method is also able to generate definitions deemed acceptable by expert evaluators, but these scored worse than human-authored definitions. Notably, evaluators with the highest level of confidence in a domain were better able to discern flaws in AI-generated definitions. We also demonstrated the ability of DRAGON-AI to incorporate natural language instructions in the form of GitHub issues. CONCLUSIONS: These findings suggest DRAGON-AIs potential to substantially aid the manual ontology construction process. However, our results also underscore the importance of having expert curators and ontology editors drive the ontology generation process.

ObjectiveAlthough supervised machine learning is popular for information extraction from clinical notes, creating large annotated datasets requires extensive domain expertise and is time-consuming. Meanwhile, large language models (LLMs) have demonstrated promising transfer learning capability. In this study, we explored whether recent LLMs could reduce the need for large-scale data annotations.Materials and methodsWe curated a dataset of 769 breast cancer pathology reports, manually labeled with 12 categories, to compare zero-shot classification capability of the following LLMs: GPT-4, GPT-3.5, Starling, and ClinicalCamel, with task-specific supervised classification performance of 3 models: random forests, long short-term memory networks with attention (LSTM-Att), and the UCSF-BERT model.ResultsAcross all 12 tasks, the GPT-4 model performed either significantly better than or as well as the best supervised model, LSTM-Att (average macro F1-score of 0.86 vs 0.75), with advantage on tasks with high label imbalance. Other LLMs demonstrated poor performance. Frequent GPT-4 error categories included incorrect inferences from multiple samples and from history, and complex task design, and several LSTM-Att errors were related to poor generalization to the test set.DiscussionOn tasks where large annotated datasets cannot be easily collected, LLMs can reduce the burden of data labeling. However, if the use of LLMs is prohibitive, the use of simpler models with large annotated datasets can provide comparable results.ConclusionsGPT-4 demonstrated the potential to speed up the execution of clinical NLP studies by reducing the need for large annotated datasets. This may increase the utilization of NLP-based variables and outcomes in clinical studies.

BackgroundPatients find technology tools to be more approachable for seeking sensitive health-related information, such as reproductive health information. The inventive conversational ability of artificial intelligence (AI) chatbots, such as ChatGPT (OpenAI Inc), offers a potential means for patients to effectively locate answers to their health-related questions digitally.ObjectiveA pilot study was conducted to compare the novel ChatGPT with the existing Google Search technology for their ability to offer accurate, effective, and current information regarding proceeding action after missing a dose of oral contraceptive pill.MethodsA sequence of 11 questions, mimicking a patient inquiring about the action to take after missing a dose of an oral contraceptive pill, were input into ChatGPT as a cascade, given the conversational ability of ChatGPT. The questions were input into 4 different ChatGPT accounts, with the account holders being of various demographics, to evaluate potential differences and biases in the responses given to different account holders. The leading question, "what should I do if I missed a day of my oral contraception birth control?" alone was then input into Google Search, given its nonconversational nature. The results from the ChatGPT questions and the Google Search results for the leading question were evaluated on their readability, accuracy, and effective delivery of information.ResultsThe ChatGPT results were determined to be at an overall higher-grade reading level, with a longer reading duration, less accurate, less current, and with a less effective delivery of information. In contrast, the Google Search resulting answer box and snippets were at a lower-grade reading level, shorter reading duration, more current, able to reference the origin of the information (transparent), and provided the information in various formats in addition to text.ConclusionsChatGPT has room for improvement in accuracy, transparency, recency, and reliability before it can equitably be implemented into health care information delivery and provide the potential benefits it poses. However, AI may be used as a tool for providers to educate their patients in preferred, creative, and efficient ways, such as using AI to generate accessible short educational videos from health care provider-vetted information. Larger studies representing a diverse group of users are needed.

IMPORTANCE: Electronic health record textual sources such as medication signeturs (sigs) contain valuable information that is not always available in structured form. Commonly processed through manual annotation, this repetitive and time-consuming task could be fully automated using large language models (LLMs). While most sigs include simple instructions, some include complex patterns. OBJECTIVES: We aimed to compare the performance of GPT-3.5 and GPT-4 with smaller fine-tuned models (ClinicalBERT, BlueBERT) in extracting the average daily dose of 2 immunomodulating medications with frequent complex sigs: hydroxychloroquine, and prednisone. METHODS: Using manually annotated sigs as the gold standard, we compared the performance of these models in 702 hydroxychloroquine and 22 104 prednisone prescriptions. RESULTS: GPT-4 vastly outperformed all other models for this task at any level of in-context learning. With 100 in-context examples, the model correctly annotates 94% of hydroxychloroquine and 95% of prednisone sigs to within 1 significant digit. Error analysis conducted by 2 additional manual annotators on annotator-model disagreements suggests that the vast majority of disagreements are model errors. Many model errors relate to ambiguous sigs on which there was also frequent annotator disagreement. DISCUSSION: Paired with minimal manual annotation, GPT-4 achieved excellent performance for language regression of complex medication sigs and vastly outperforms GPT-3.5, ClinicalBERT, and BlueBERT. However, the number of in-context examples needed to reach maximum performance was similar to GPT-3.5. CONCLUSION: LLMs show great potential to rapidly extract structured data from sigs in no-code fashion for clinical and research applications.

Abstract: With the advent of ChatGPT and other large language models, the number of artificial texts we encounter on a daily basis is about to increase substantially. This essay asks how this new textual situation may influence what one can call the “standard expectation of unknown texts,” which has always included the assumption that any text is the work of a human being. As more and more artificial writing begins to circulate, the essay argues, this standard expectation will shift—first, from the immediate assumption of human authorship to, second, a creeping doubt: did a machine write this? In the wake of what Matthew Kirschenbaum has called the “textpocalypse,” however, this state cannot be permanent. The author suggests that after this second transitional period, one may suspend the question of origins and, third, take on a post-artificial stance. One would then focus only on what a text says, not on who wrote it; post-artificial writing would be read with an agnostic attitude about its origins. This essay explores the implications of such post-artificiality by looking back to the early days of text synthesis, considering the limitations of aesthetic Turing tests, and indulging in reasoned speculation about the future of literary and nonliterary text generation.

PurposeReal world evidence is crucial to understanding the diffusion of new oncologic therapies, monitoring cancer outcomes, and detecting unexpected toxicities. In practice, real world evidence is challenging to collect rapidly and comprehensively, often requiring expensive and time-consuming manual case-finding and annotation of clinical text. In this Review, we summarise recent developments in the use of artificial intelligence to collect and analyze real world evidence in oncology.MethodsWe performed a narrative review of the major current trends and recent literature in artificial intelligence applications in oncology.ResultsArtificial intelligence (AI) approaches are increasingly used to efficiently phenotype patients and tumors at large scale. These tools also may provide novel biological insights and improve risk prediction through multimodal integration of radiographic, pathological, and genomic datasets. Custom language processing pipelines and large language models hold great promise for clinical prediction and phenotyping.ConclusionsDespite rapid advances, continued progress in computation, generalizability, interpretability, and reliability as well as prospective validation are needed to integrate AI approaches into routine clinical care and real-time monitoring of novel therapies.

OBJECTIVE: Vaccination has engendered a spectrum of public opinions, with social media acting as a crucial platform for health-related discussions. The emergence of artificial intelligence technologies, such as large language models (LLMs), offers a novel opportunity to efficiently investigate public discourses. This research assesses the accuracy of ChatGPT, a widely used and freely available service built upon an LLM, for sentiment analysis to discern different stances toward Human Papillomavirus (HPV) vaccination. METHODS: Messages related to HPV vaccination were collected from social media supporting different message formats: Facebook (long format) and Twitter (short format). A selection of 1,000 human-evaluated messages was input into the LLM, which generated multiple response instances containing its classification results. Accuracy was measured for each message as the level of concurrence between human and machine decisions, ranging between 0 and 1. RESULTS: Average accuracy was notably high when 20 response instances were used to determine the machine decision of each message: .882 (SE = .021) and .750 (SE = .029) for anti- and pro-vaccination long-form; .773 (SE = .027) and .723 (SE = .029) for anti- and pro-vaccination short-form, respectively. Using only three or even one instance did not lead to a severe decrease in accuracy. However, for long-form messages, the language model exhibited significantly lower accuracy in categorizing pro-vaccination messages than anti-vaccination ones. CONCLUSIONS: ChatGPT shows potential in analyzing public opinions on HPV vaccination using social media content. However, understanding the characteristics and limitations of a language model within specific public health contexts remains imperative.

BACKGROUND: Amyloidosis, a rare multisystem condition, often requires complex, multidisciplinary care. Its low prevalence underscores the importance of efforts to ensure the availability of high-quality patient education materials for better outcomes. ChatGPT (OpenAI) is a large language model powered by artificial intelligence that offers a potential avenue for disseminating accurate, reliable, and accessible educational resources for both patients and providers. Its user-friendly interface, engaging conversational responses, and the capability for users to ask follow-up questions make it a promising future tool in delivering accurate and tailored information to patients. OBJECTIVE: We performed a multidisciplinary assessment of the accuracy, reproducibility, and readability of ChatGPT in answering questions related to amyloidosis. METHODS: In total, 98 amyloidosis questions related to cardiology, gastroenterology, and neurology were curated from medical societies, institutions, and amyloidosis Facebook support groups and inputted into ChatGPT-3.5 and ChatGPT-4. Cardiology- and gastroenterology-related responses were independently graded by a board-certified cardiologist and gastroenterologist, respectively, who specialize in amyloidosis. These 2 reviewers (RG and DCK) also graded general questions for which disagreements were resolved with discussion. Neurology-related responses were graded by a board-certified neurologist (AAH) who specializes in amyloidosis. Reviewers used the following grading scale: (1) comprehensive, (2) correct but inadequate, (3) some correct and some incorrect, and (4) completely incorrect. Questions were stratified by categories for further analysis. Reproducibility was assessed by inputting each question twice into each model. The readability of ChatGPT-4 responses was also evaluated using the Textstat library in Python (Python Software Foundation) and the Textstat readability package in R software (R Foundation for Statistical Computing). RESULTS: ChatGPT-4 (n=98) provided 93 (95%) responses with accurate information, and 82 (84%) were comprehensive. ChatGPT-3.5 (n=83) provided 74 (89%) responses with accurate information, and 66 (79%) were comprehensive. When examined by question category, ChatGTP-4 and ChatGPT-3.5 provided 53 (95%) and 48 (86%) comprehensive responses, respectively, to general questions (n=56). When examined by subject, ChatGPT-4 and ChatGPT-3.5 performed best in response to cardiology questions (n=12) with both models producing 10 (83%) comprehensive responses. For gastroenterology (n=15), ChatGPT-4 received comprehensive grades for 9 (60%) responses, and ChatGPT-3.5 provided 8 (53%) responses. Overall, 96 of 98 (98%) responses for ChatGPT-4 and 73 of 83 (88%) for ChatGPT-3.5 were reproducible. The readability of ChatGPT-4s responses ranged from 10th to beyond graduate US grade levels with an average of 15.5 (SD 1.9). CONCLUSIONS: Large language models are a promising tool for accurate and reliable health information for patients living with amyloidosis. However, ChatGPTs responses exceeded the American Medical Associations recommended fifth- to sixth-grade reading level. Future studies focusing on improving response accuracy and readability are warranted. Prior to widespread implementation, the technologys limitations and ethical implications must be further explored to ensure patient safety and equitable implementation.

Here, we summarize the proceedings of the inaugural Artificial Intelligence in Primary Immune Deficiencies conference, during which experts and advocates gathered to advance research into the applications of artificial intelligence (AI), machine learning, and other computational tools in the diagnosis and management of inborn errors of immunity (IEIs). The conference focused on the key themes of expediting IEI diagnoses, challenges in data collection, roles of natural language processing and large language models in interpreting electronic health records, and ethical considerations in implementation. Innovative AI-based tools trained on electronic health records and claims databases have discovered new patterns of warning signs for IEIs, facilitating faster diagnoses and enhancing patient outcomes. Challenges in training AIs persist on account of data limitations, especially in cases of rare diseases, overlapping phenotypes, and biases inherent in current data sets. Furthermore, experts highlighted the significance of ethical considerations, data protection, and the necessity for open science principles. The conference delved into regulatory frameworks, equity in access, and the imperative for collaborative efforts to overcome these obstacles and harness the transformative potential of AI. Concerted efforts to successfully integrate AI into daily clinical immunology practice are still needed.