Sheng-He Gu

Fascinating insights shared by NEJM AI on the limitations of #LLMs in medical coding! The study by Ali Soroush, MD, MS, and colleagues reveals critical benchmarks in LLMs' ability to map medical codes accurately. Key findings highlight the error-proneness of these models, with #GPT-4 showing the best but still limited performance. The in-depth analysis provided in the study underscores the complexities of medical code querying and the need for extensive training and perhaps model adjustments to ensure precision and reliability. The real takeaway? There's a significant journey ahead in harnessing AI effectively in healthcare administrative tasks. That is why there is a need for medical-specific models trained on medical knowledge and not just random stuff on the internet. 🔗 https://lnkd.in/g2XtCYPC #AIinMedicine #HealthTech #ArtificialIntelligence #MedicalCoding HOPPR

16

2 Comments

In computational linguistics, researchers are focused on enhancing language models' ability to extract precise information from extensive textual data. Recent studies introduce "retrieval heads," specialized attention mechanisms that selectively focus on critical segments of large texts, improving targeted data retrieval. These heads, integrated into transformer-based models like LLaMA and Yi, significantly enhance accuracy and efficiency, particularly in handling long-context scenarios. The methodology involves thorough experiments across various models, including the Needle-in-a-Haystack test, which embeds specific information to measure precision. Results show models with active retrieval heads outperform those without, with accuracy dropping notably when these heads are masked. The empirical evidence underscores the effectiveness of retrieval heads in improving information retrieval precision and reliability. This research deepens our understanding of attention mechanisms in large-scale text processing, paving the way for more efficient and accurate language models. https://lnkd.in/dNNPvMXw https://lnkd.in/ddbHX6EA Innovation Hacks AI #aiinnovation #generativeai

1

I have my first interactive text with bioinformatic questions of defining biological questions for computational algorithms for finding the origin of replication in bacterial genome.

1

Cracking the Code: Reverse and complement nucleic acid sequences (DNA, RNA)

1

𝗟𝗟𝗠 𝗔𝗴𝗲𝗻𝘁𝘀 - 𝗧𝗵𝗲 𝗡𝗲𝘅𝘁 𝗘𝘃𝗼𝗹𝘂𝘁𝗶𝗼𝗻 𝗼𝗳 𝗟𝗟𝗠𝘀 LLM Agents, also known as Agent AI Workflows, represent the next major evolution in Large Language Models (LLMs) like ChatGPT. These agents significantly enhance the capability of LLMs to handle complex queries, manage private data efficiently, and reduce operational costs associated with using LLM technologies. 🔍 𝗖𝘂𝗿𝗿𝗲𝗻𝘁 𝗖𝗵𝗮𝗹𝗹𝗲𝗻𝗴𝗲𝘀 𝘄𝗶𝘁𝗵 𝗟𝗟𝗠𝘀: While LLMs perform exceptionally well with public data, they face limitations when users attempt to integrate their own large-scale data, primarily due to the token constraints of models like GPT-4, which can handle a maximum of 8k tokens. Traditional methods such as Retrieval Augmented Generation (RAG) fall short, especially when the data necessary to answer questions doesn't directly match the query terms. A simple question like "Find me patients who are teenagers" is very hard to answer using current methods. 🔗 𝗛𝗼𝘄 𝗟𝗟𝗠 𝗔𝗴𝗲𝗻𝘁𝘀 𝗦𝗼𝗹𝘃𝗲 𝗧𝗵𝗲𝘀𝗲: LLM Agents are designed to overcome these constraints by inversely integrating "tools" that an LLM can leverage to fetch relevant data without the need for direct data embedding in the prompt. These tools allow LLMs to dynamically retrieve data relevant to the queries posed, enabling responses to complex questions that would typically not be directly referenced in the data. 🛠️ 𝗛𝗼𝘄 𝗟𝗟𝗠 𝗔𝗴𝗲𝗻𝘁𝘀 𝗪𝗼𝗿𝗸: LLM Agents employ a multi-step workflow where each query activates a specific tool based on the requirement of the question. This modular approach enables the LLM to process queries that involve multiple data points and sources, mimicking real-world problem-solving methods more closely than traditional single-step data retrieval. 💡𝗔𝗱𝘃𝗮𝗻𝘁𝗮𝗴𝗲𝘀 𝗼𝗳 𝗟𝗟𝗠 𝗔𝗴𝗲𝗻𝘁𝘀: 1. 𝘋𝘢𝘵𝘢 𝘌𝘧𝘧𝘪𝘤𝘪𝘦𝘯𝘤𝘺: By minimizing the data directly processed by the LLM, costs and risks associated with private data are significantly reduced. 2. 𝘈𝘯𝘴𝘸𝘦𝘳 𝘊𝘰𝘮𝘱𝘭𝘦𝘹 𝘘𝘶𝘦𝘴𝘵𝘪𝘰𝘯𝘴: Agents can answer questions that do have the same words as the the content used to answer them. Agents can also a multi-step process to leverage multiple data sources. 3. 𝘐𝘮𝘱𝘳𝘰𝘷𝘦𝘥 𝘈𝘤𝘤𝘶𝘳𝘢𝘤𝘺 𝘢𝘯𝘥 𝘙𝘦𝘭𝘪𝘢𝘣𝘪𝘭𝘪𝘵𝘺: The ability to cross-reference and validate answers through multiple data points reduces the risk of inaccuracies and "hallucinations" common in traditional LLM outputs. Read this article to learn more about LLM Agents and how to start using them today... How are you providing your private data to the LLM today?

18

3 Comments

Below webinar shares information about very interesting non-fixing sample preservation solution for short-term storage of tissue, cell or blood samples for any downstream analysis including Single Cell Multiomics using BD Rhapsody workflow. Register below. #OMG #singlecellanalysis #omics #cellularanalysis

6

I just came across a fascinating article by Casey Sackett, Devin Harper, and Aaron Pavez from Webot in IEEE Spectrum Magazine about the exciting (and tricky!) world of large language models (LLMs) in mental health. In essence, they describe that current LLMs are challenging to control for applications for which these models are not trained and might not behave as wished. One example is that the data and objective the LLM are trained on is to provide answers and not help develop insights which could lead to a change in perception and behavior. An even more significant challenge is that uncontrolled use of LLMs could create serious risks for people seeking help in the healthcare domain. This has the effect that these models shy away from such situations and bail on responses. Nevertheless, the article is optimistic (as I am too) and outlines that the technology can be used and integrated into more traditional frameworks. Although (pre-trained) LLMs have many issues, there is still much potential left to improve them fundamentally. If you're curious about the future of AI in healthcare, this 11-minute read is worth your time! ⏱️ https://lnkd.in/exVqep2z

19

I’m thrilled to share our latest research on predictive analytics in healthcare! 🏥📊 This study introduces a predictive framework for ICU length of stay (LOS) using hospital EHR data and machine learning models. Utilizing explainable AI (xAI), it achieves 98% AUC with XGBoost, enhancing hospital efficiency and patient outcomes by providing trustworthy LOS predictions and explaining model outputs. Belal Alsinglawi Omar Mubin #HealthcareInnovation #MachineLearning #PredictiveAnalytics #AIHospitalManagement #UAEU

23

Today, I'll discuss the crucial role of caution in developing and deploying modern AI applications in healthcare at the AI for Business event at Frank G. Zarb School of Business. On one hand, AI can transform healthcare accessibility by reducing wait times, assisting doctors in capturing patient histories and understanding complex conditions more accurately and quickly, relieving nurses and doctors of menial tasks to focus on patient care, accelerating drug and vaccination discovery, and effectively combating future pandemics. On the other hand, these systems can undermine our civil liberties, exacerbate social and systemic biases, and violate trust. I'll explore some developments that illustrate both impacts.

11

While the consumers like you & me are lured into the speed, efficacy and magic of #GenAI, the worried #AI developers are looking for ways to understand the inner working of LLMs for future safety. While #interpretability is one way to address it (Anthropic showcased it earlier), #explainability, especially by #LLM itself (since they are good at producing prose or poetry), could be another. The latest paper published a week ago explains #SaySelf, a training framework that attempts to addresses this requirement by improving LLM confidence estimates. It uses multiple sampled reasoning chains, semantic clustering and #RL with a newly crafted custom-reward function to calibrate confidence estimates and finetune an LLM. SaySelf reduces #confidence #calibration error by 30%. Link to the paper & code in comments.

17

1 Comment

We agree! We are focusing on this use case at One Fact Foundation - if anyone wants to help (or needs help!) please reach out: hello@onefact.org or DM me. We have helped several health systems integrate my work on AI from grad school (https://lnkd.in/e-DACRwr) which is at 800+ citations now, and used this system in everything from: - improving recapture rates for the Centers for Medicare & Medicaid Services' hierarchical condition category codes. This can lead to increased reimbursement from government Medicare Advantage programs prevalent in Accountable Care Organizations or Ambulatory Surgical Centers - especially for private equity-backed care orgs :) - improving the prediction of diagnosis codes (usually the World Health Organization's international classification codes that rockstars like Tahilia J. Rebello, PhD help develop by coordinating thousands of clinicians worldwide) - reducing the burden on clinicians to take notes rather than spend time with patients. Often note-taking happens on paper, and in the United States, is bound by legal liability and cash incentives, or cash-related incentives like relative value units used to improve coding integrity in revenue cycle management programmes and increase case-mix index and reimbursement. Happy to chat about any and all of the above, especially for under-served (rural, remote, poor, low- and middle-income country tertiary care) systems. A dream would be to support the ASHA Aide programme in Ministry of Health and Family Welfare, Government of India or the Hukou system run by the Ministry of Health of the People's Republic of China. If you know anyone, feel free to cc me (jaan@onefact.org). Thanks to Theresa Osborne, Yubin Park, PhD, Christopher Whaley, @cynthia fisher (Power To The Patients and Patient Rights Advocate), Brown Institute for Media Innovation, the AIM-AHEAD Consortium, Michela Magas, Steven C. Hayes, Philip Resnik and all of our other funders and supporters. This would not be possible without your support! Excited to drive the cost of health care delivery as close to zero as possible.

9

Join us on June 5, 2024, at 1 PM (ET) for a webinar on tackling the reproducibility and generalizability challenges in #MachineLearning ! Learn strategies for data accessibility, leakage prevention, preprocessing, algorithm selection, and more to ensure reliable ML outcomes. Farhad Maleki Society for Imaging Informatics in Medicine (SIIM) Register: ⬇ https://lnkd.in/dqrDUg2J Objectives: - Equip participants with strategies to facilitate data accessibility, prevent data leakage, and standardize data preprocessing to enhance the reproducibility of machine learning projects. - Provide insights into selecting appropriate algorithms, promoting code sharing, and managing dependencies effectively to achieve reliable machine learning outcomes. - Advocate for the adoption of comprehensive and transparent research documentation standards to improve the clarity and reproducibility of machine learning research. - Demonstrate how to utilize advanced tools and frameworks and adhere to community standards to significantly improve the reproducibility of machine learning systems. #ML #DataScience #Healthcare #Webinar #AI

14

"With large language models that take notes during patient visits and algorithms that identify disease, artificial intelligence has begun to prove its worth as an assistant for physicians. But a new study from Stanford Medicine shows the potential of AI as a facilitator—one that helps doctors and nurses connect to achieve more efficient, effective patient care. The study, which was published in JAMA Internal Medicine, describes an AI-based model in use at Stanford Hospital that predicts when a patient is declining and flags the patient's physicians and nurses." #ai #medicalcare #doctors #nurses #collaboration #communications

1

1 Comment

Using transformers to pretrain a foundational genomic model.

2