NEJM AI’s Post

This is a fantastic application of AI technologies and I’m excited to see this direction of holistic patterns for eligibility assessment. Following the cautionary discussion simply reinforces the case for human/physician in the loop architectures for clinical AI. Can we also take a moment to acknowledge the impact of NEJM AI as a gathering place for clinical and medicinal AI research? Such an exciting time in the field. #nejm #nejmai #ai #llm #rag

A recent study led by the University of Cambridge found that GPT-4, a large language model, has clinical knowledge and reasoning skills approaching those of specialist eye doctors. GPT-4 scored significantly better than unspecialized junior doctors in a test measuring specialist eye knowledge. GPT-4 gained similar scores to trainee and expert eye doctors, with the exception of the top performing doctors who scored higher. While the researchers don't believe that large language models will replace healthcare professionals, they do see the potential for models like GPT-4 to improve healthcare as part of the clinical workflow. For example, these models could be used to provide eye-related advice, diagnosis, and management suggestions in well-controlled contexts, such as triaging patients or where access to specialist healthcare professionals is limited. This study highlights the potential for large language models to contribute to healthcare. Interesting times ahead of us how this technology will continue to evolve and improve patient outcomes. #healthcare #technology #artificialintelligence

GPT-4, a large language model, was recently tested against doctors of varying expertise levels in diagnosing and treating eye problems. The results were impressive, with GPT-4 outscoring unspecialized junior doctors, who have a similar level of specialist eye knowledge to general practitioners. Additionally, GPT-4 achieved similar scores to trainee and expert eye doctors. While the top-performing doctors scored higher, the researchers noted that large language models have the potential to improve healthcare as part of the clinical workflow. It's important to note that these models aren't intended to replace healthcare professionals, but rather work alongside them to enhance patient care. Read more about this study here: https://lnkd.in/dxAwN5YS

"...Strategies to mitigate bias in language models are a very active research area, commonly divided into approaches that augment or alter the model training data, strategies that enhance or provide additional model training, strategies that alter the model execution, and strategies that apply corrections post-processing.6 Supplementing such models with additional training through data augmentation or additional reinforcement learning from human feedback might help to some extent, but are unlikely to be completely successful as it is largely outside of the potential of the underlying technology to avoid bias completely.." "...Advances in language technology offer many exciting possibilities for medicine, including possibilities to reduce the time-consuming burden of documentation9 and thereby free more clinical time for human-focused work. In such workflows, biases might still have an effect that should be considered, but the overall net benefit will be enormous, and the risk mitigation might be more straightforward than for other applications of the technology. However, for tasks such as clinical decision making and treatment planning, the risks are potentially severe and there is a substantial requirement for diligent implementation strategies to avoid repeating algorithmic harms such as those that have been made in the past..." Janna Hastings https://lnkd.in/dU_9cbfk

🌟 Unveiling the Potential of Large Language Models in Healthcare with Stanford University HAI 🌟 Fei-Fei Li 🔬 As we continue to explore the integration of large language models (LLMs) into healthcare, it's crucial to address the current state of these technologies. Despite their promising capabilities, there's still a significant journey ahead to fully harness their potential in clinical settings. Stanford's Human-Centered AI Institute (Stanford HAI) is at the forefront of this exploration, emphasizing the need for a robust, systematic evaluation process. 🧑⚕️💡 📈 Recent studies, including those highlighted by Stanford HAI, have pinpointed the challenges and limitations of LLMs in healthcare. One concerning finding from a STATnews report was the occurrence of safety errors in LLM responses to patient messages, with some advice potentially being harmful. This underscores the gap between LLMs' theoretical promise and their practical utility in real-world medical environments. 🏥🖥️ 🤖 Stanford HAI advocates for an increased focus on testing LLMs using real patient data and evaluating them across a spectrum of medical tasks from diagnostics to administrative support, which could significantly reduce physician burnout. This approach aims to refine the algorithms before they are deployed, ensuring safety, fairness, and efficiency. 📊🔍 🌍 The goal is to create a future where generative AI can support healthcare professionals by improving workflow efficiency and patient outcomes without compromising safety. Stanford HAI's ongoing research and initiatives strive to set the standards for how these technologies should be evaluated and implemented. 🌐✨ https://lnkd.in/eKFPDnNV

Large language models (LLMs) can potentially democratize access to medical knowledge. While many efforts have been made to harness and improve LLMs' medical knowledge and reasoning capacities, the resulting models are either closed-source (e.g., PaLM, GPT-4) or limited in scale (<= 13B parameters), which restricts their abilities. In this work, we improve access to large-scale medical LLMs by releasing MEDITRON: a suite of open-source LLMs with 7B and 70B parameters adapted to the medical domain. MEDITRON builds on Llama-2 (through our adaptation of Nvidia's Megatron-LM distributed trainer), and extends pretraining on a comprehensively curated medical corpus, including selected PubMed articles, abstracts, and internationally-recognized medical guidelines. Evaluations using four major medical benchmarks show significant performance gains over several state-of-the-art baselines before and after task-specific finetuning. Overall, MEDITRON achieves a 6% absolute performance gain over the best public baseline in its parameter class and 3% over the strongest baseline we finetuned from Llama-2. Compared to closed-source LLMs, MEDITRON-70B outperforms GPT-3.5 and Med-PaLM and is within 5% of GPT-4 and 10% of Med-PaLM-2.

This paper explores the integration of Generative Large Language Models (LLMs) in clinical settings to enhance evidence-based medicine (EBM). 1️⃣ The study evaluates the capability of LLMs like ChatGPT-3.5, GPT-4, and others in autonomously operating in clinical settings, focusing on their effectiveness in ordering relevant investigations and adhering to clinical guidelines. 2️⃣ LLMs were tested against real-world clinical scenarios, with their performance assessed based on correctness, judicious use of tools, guideline conformity, and resistance to hallucinations. 3️⃣ GPT-4 demonstrated superior performance across various specialties and difficulty levels, showing effective integration with clinical tooling and adherence to guidelines. 4️⃣ The study highlights limitations such as model's ability to handle complex guidelines, diagnostic nuances, and the risk of generating incorrect or fabricated information (hallucinations). 5️⃣ Retrieval Augmented Generation (RAG) was used to tailor recommendations more specifically to patients and healthcare systems. 6️⃣ The paper suggests that prompt engineering and LLM reasoning can significantly contribute to healthcare, transforming patient management and clinical decision-making. The paper is worth reading as it presents evidence of the potential of LLMs like GPT-4 in autonomously performing tasks in clinical care, demonstrating their ability to interact with real-world healthcare systems and manage patient care in a guideline-directed manner. ✍🏻 Akhil Vaid, Joshua Lampert, Juhee Lee, Ashwin Sawant, Donald Apakama, Ankit Sakhuja, Ali Soroush, Denise Lee, Isotta Landi, Nicole Bussola, Ismail Nabeel, Robbie Freeman, Patricia Kovatch, Brendan Carr, Benjamin Glicksberg, Edgar Argulian, Stamatios Lerakis, Monica Kraft, Alexander Charney, and Girish Nadkarni. Generative Large Language Models are autonomous practitioners of evidence-based medicine. arXiv, 2024, 2401.02851. [Not peer-reviewed] ✅ Subscribe to my newsletter and stay at the forefront of groundbreaking studies. Get started here: https://lnkd.in/eR7qichj.

As the potential for AI in healthcare grows, it's critical to ensure that large language models (LLMs) can effectively handle real-world clinical tasks. However, evaluating their performance has been challenging due to a lack of comprehensive benchmarks that reflect the complexity and diversity of healthcare settings. Three recent papers present novel benchmarks that aim to fill this gap and accelerate the development and integration of LLMs into clinical practice. 1️⃣ ‘Large Language Models in Healthcare: A Comprehensive Benchmark’ presents BenchHealth, a benchmark for medical reasoning, generation, and understanding, includes seven tasks and 13 datasets from various healthcare domains. The evaluation covers five metrics: matching, faithfulness, comprehensiveness, generalizability, and robustness. The results show that LLMs perform well in closed-ended QA tasks, with GPT-4 achieving 95% accuracy on MedQA, but struggle with open-ended questions and language comprehension, with scores as low as 40%. Full text: https://lnkd.in/d7MXyi4v 2️⃣- ‘LongHealth: A Question Answering Benchmark with Long Clinical Documents’: This benchmark evaluates LLMs’ ability to process long clinical documents, covering 20 patient cases ranging from 5,090 to 6,754 words per case. The benchmark includes 400 multiple-choice questions testing information extraction, negation handling, and temporal reasoning. Mixtral-8x7B-Instruct-v0.1 achieved the highest accuracy at 77%, while all models faced challenges with identifying missing information, with accuracies as low as 23%. Full text: https://lnkd.in/dZYS-Ax2 3️⃣ ‘MedAlign: A Clinician-Generated Dataset for Instruction Following with Electronic Medical Records’ presents a dataset of 983 clinician-generated instructions for EHR-based tasks. It enables the evaluation of LLMs for complex summarization, care planning, and diagnostic support based on real-world longitudinal patient records. The evaluation revealed significant gaps, with accuracies ranging from 29% to 65%, highlighting the need for further model refinement. Full text: https://lnkd.in/dbGRvS8K These benchmarks offer valuable insights into the current state of LLMs in healthcare and highlight the need for collaboration between clinicians, researchers, and NLP experts. By creating datasets reflecting real-world clinical workflows and challenges, we can develop more reliable AI tools to support healthcare professionals and improve patient care.

This paper investigates the synergistic potential of Large Language Models (LLMs) and biomedical knowledge graphs in training semantic models for biomedical and clinical applications. These semantic models, adept at interpreting meaning and context rather than just processing data, can discern nuanced language use, such as understanding "feeling blue" as an expression of sadness. A key focus of this paper is addressing the scarcity of human-written definitions in biomedical thesauri. It demonstrates the effectiveness of LLMs in generating comprehensive definitions for a majority of biomedical concepts. The fusion of knowledge graph insights with LLMs' generalization capabilities is poised to improve healthcare applications, including clinical summarization, decision support, patient diagnosis, pharmacovigilance, and drug discovery. This approach also supports the creation of more interpretable AI systems in healthcare. #insightsonai.com https://lnkd.in/gzDEuHzK

This study explores enhancing the accuracy of patient summaries generated by large language models by focusing on minimizing hallucinations. It introduces a novel dataset and annotation protocol to identify and mitigate hallucinations, ultimately improving the reliability of generated patient summaries. 1️⃣ Developed a new dataset, MIMIC-IV-Note-DI, for training LLMs to produce more accurate patient summaries, with medical experts involved in the annotation process to ensure data quality. 2️⃣ Demonstrated the efficacy of fine-tuning LLMs on a curated dataset to significantly reduce the occurrence of hallucinations, thus enhancing the faithfulness of the summaries. 3️⃣ GPT-4 showed promising results in generating high-quality patient summaries with fewer hallucinations and in automatic hallucination detection. 4️⃣ Common quantitative metrics did not correlate well with the faithfulness and quality of the generated summaries, highlighting the need for qualitative evaluations. The paper provides valuable insights into the use of LLMs in healthcare, showing that careful dataset preparation and evaluation can lead to the generation of more reliable patient summaries, a crucial step towards their practical application in medical settings. ✍🏻 Stefan Hegselmann, Shannon Zejiang Shen, Florian Gierse, Monica Agrawal, David Sontag, Xiaoyi Jiang. A Data-Centric Approach To Generate Faithful and High Quality Patient Summaries with Large Language Models. arXiv. 23 Feb 2024. arXiv:2402.15422 ✅ Join +28k other readers and subscribe to our monthly newsletter: https://lnkd.in/eR7qichj.