Interesting reads ... May 2024
Zeljko K. , Dan Bean , Anthony Shek, PhD , Rebecca Bendayan, Ph.D. , Harry Hemingway FFPH FRCP FMedSci , Josh Au Yeung , Alexander T Deng , Alfred Baston, Jack Ross, Esther Idowu , James T Teo, and Richard J B Dobson evaluated a generative transformer model that integrates free text and structured EHR data to predict various future medical outcomes. Trained on datasets from three hospitals involving over 800,000 patients, the model demonstrated strong performance metrics and offers a novel tool for clinical research and education but is not yet suitable for direct clinical decision support.
Clare McGenity , Emily Clarke , Charlotte Jennings , Gillian Matthews, Caroline Cartlidge , Henschel Freduah-Agyemang , Deborah Stocken , and Darren Treanor found that AI models in digital pathology demonstrated high diagnostic accuracy, with a mean sensitivity of 96.3% and specificity of 93.3%. Despite this, they noted significant variability in study designs and a high risk of bias, underscoring the need for more rigorous evaluation of AI performance in this field.
Prof. Andreas Charalambous identifies performance and effort expectancy, regulatory challenges, skepticism towards technology, and organizational barriers as primary obstacles to digital transformation in healthcare, proposing governance frameworks like TAPIC to guide successful integration. Despite these challenges, digital transformation can enhance patient care, prevent adverse drug events, and reduce healthcare costs, with recognition from the WHO and European Commission on its potential to improve access, quality, and efficiency in healthcare delivery.
Cathy Ong Ly , Balagopal Unnikrishnan , Tony Tadic , Tirth Patel, Joe Duhamel, Sonja Kandel , Yasbanoo Moayedi, Michael Brudno , Andrew Hope, Dr. Heather Ross , and Chris McIntosh found that shortcut learning in medical AI, where models rely on spurious correlations, leads to significant overestimation of performance by up to 20%. They propose a new method for estimating external accuracy by calibrating for data acquisition biases, which improves accuracy estimates to within 4%, as demonstrated across 13 datasets including various medical data types.
Hartmut H?ntze , Lina Xu , Felix Dorfner ,?Leonhard Donle, Daniel Truhn ,?Hugo Aerts, Mathias Prokop , Bram van Ginneken , Alessa Hering , Lisa Adams , and Keno Bressem present MRSegmentator, a deep learning model for multi-organ segmentation in MRI and CT scans that overcomes MRI-specific challenges like resolution and intensity variability. Trained on a large, diverse dataset, the model achieves high segmentation accuracy, with Dice Similarity Coefficients of up to 0.97 for lungs and heart, and is available as an open-source resource for the medical imaging community.
Ali Soroush, MD, MS , Ben Glicksberg , Eyal Zimlichman, MD , Yiftach Barash , Robert Freeman, Alexander Charney , Girish Nadkarni , and Eyal Klang evaluated the performance of various LLMs in generating medical billing codes, finding significant inaccuracies across models. GPT-4 performed the best with exact match rates up to 49.8%, yet still demonstrated substantial error rates, underscoring the limitations and potential risks of current LLM applications in medical coding.
领英推荐
Weiwei Huo, Xinze Yuan , Xianmiao Li, Wenhao Luo, Jiaying Xie, and Bowen Shi explore how user participation influences the acceptance of medical AI among healthcare professionals. Their findings reveal that active involvement in AI development boosts medical staff's acceptance through improved self-efficacy and reduced anxiety, highlighting the dual roles of staff as both service providers and AI users.
Stuart McLennan , Amelia Fiske , Leo Anthony Celi conducted a study involving intensive care professionals from 24 countries, finding general enthusiasm about AI's potential to enhance patient outcomes and manage data overload in ICUs, but significant barriers exist. Key issues include the lack of digital infrastructure, particularly outside large tertiary hospitals in high-income countries, and a notable skills gap among ICU staff, underscoring the need for foundational digital infrastructure, comprehensive training, and closer collaboration between clinicians and AI developers.
Hong-Yu Zhou, Subathra Adithan , Julián Nicolás Acosta , Eric Topol, MD , and Pranav Rajpurkar present MedVersa, an AI model leveraging both visual and linguistic inputs to excel in multifaceted medical imaging tasks, outperforming specialized models. Supported by the extensive MedInterp dataset, MedVersa demonstrates state-of-the-art performance across nine tasks and offers adaptability in real-world clinical settings, enhancing diagnostic efficiency and addressing biases in medical AI.
Kevin Wu, Eric Wu, Ally Casasola, Angela Zhang, Kevin Wei, Teresa Nguyen, Sith Riantawan, Patricia Shi, Daniel Ho, and James Zou found that 50-90% of medical responses generated by LLMs were not fully supported by their cited sources, with GPT-4 still having 30% unsupported individual statements despite using retrieval-augmented generation. They also observed that many URLs generated by LLMs were invalid, contributing to citation issues, and noted significant variability in the quality of sources provided, with response-level support ranging from 7% to 54% across different models.
VPO at Telegenisys
5 个月Spoilers: These future health prediction systems will be ever more accurate as an implementation of Psychohistory (the fictional kind from Asimov's Foundation series) and worse than useless as a death clock on the individual. And just as in the Foundation series each and every black swan event will "flush the cache" and make the model useless until retrained with the new facts on the ground. (Oh, mass respiratory illness as predicted years ago! Well that changes everything.) So how does the machine determine when it's "kung-fu no good here" and cash out the future bets on sub-prime lives before the market crashes, or will this gut animal response always be the responsibility of a human who has been carefully trained to apply common sense?
ICT CONSULTANT
5 个月Thank you Jan
Autodidacte ? Chargé d'intelligence économique ? AI hobbyist ethicist - ISO42001 ? Polymathe ? éditorialiste & Veille stratégique - Times of AI ? Techno-optimiste ?
5 个月Thanks so much Jan ??
Thanks so much Jan ??
Doctor of Financial technology & AI in Accounting
5 个月Thanks for sharing, "Exciting to see such a comprehensive list of readings on Digital Health & Innovation for May 2024. The intersection of technology and healthcare is crucial for advancing patient care and medical research. Looking forward to diving into these papers and discussing the innovative ideas they present!"