Meta 用 Muse 模型加速健康数据科学研究，四周产出十篇医学手稿

One big reason we are building artificial intelligence at Meta is to empower people to accelerate scientific research and understanding. One domain we are particularly excited about is health science, because progress in understanding health directly changes people’s lives through earlier diagnosis, personalized treatments, and affordable care. While the data-driven nature of health science makes it appropriate for application of AI, research in health science depends on imperfect real-world data and consequential choices about cohorts, variables, methods, and interpretation. As a consequence, our understanding of how to use AI to accelerate research in health science is still relatively early.

In this initial exploration, we developed a human-led AI system for conducting health research, in which AI accelerates each stage of research while domain experts remain responsible for the scientific decisions and conclusions. We built an internal coding scaffold powered by a combination of Muse Spark 1.3 and Muse Wonder that could access both publicly available data as well as a licensed retrospective database of five million de-identified electronic health records. We recruited a cohort of medical researchers with backgrounds in cardiology, epidemiology, infectious disease, and internal medicine, who used the coding scaffold to conduct original research over the course of just four weeks. They used the Muse scaffold to brainstorm hypotheses, write and execute all code, and collaborated with the model to write the manuscripts. Throughout the process, they also developed a shared framework and set of skills to steer the model to better accelerate their work at each stage.

Ten Research Manuscripts

Today we are sharing ten research manuscripts generated throughout this process, spanning infectious disease, sports medicine, emergency medicine, oncology, endocrinology, oral health, and more. Each paper went through several levels of iteration: feedback from peers in the medical cohort, review of analysis by a biostatistician, and finally, third-party contracted independent experts who gave summary reviews, shown below each paper.

Temporal Patterns of Arrhythmia and Conduction Disorder Diagnoses After Infective Endocarditis

Paddy Ssentongo, MD, PhD^1,2; W. Frank Peacock, MD, FACEP, FACC, FESC³; Kritika Singh, PhD⁴; Jared C. Durieux, PhD, MS, MDS, MPH⁵; Christos Colovos, MD, PhD

1. ¹Department of Medicine, Division of Infectious Diseases, Penn State Health Milton S. Hershey Medical Center, Hershey, Pennsylvania
2. ²Department of Public Health Sciences, Penn State College of Medicine, Hershey, Pennsylvania
3. ³Henry JN Taub Department of Emergency Medicine, Baylor College of Medicine, Houston, Texas
4. ⁴Stanford University School of Medicine, Stanford, California
5. ⁵Evidence Architecture, Stow, Ohio

Summary

Ssentongo et al. showed that after infective endocarditis, a serious infection of the heart's inner lining and valves, patients had a 73% higher rate of new arrhythmias and conduction disorders than patients with bacteremia without infective endocarditis (bacteria in the bloodstream without infection of the heart). One-year risk was seven percent vs four percent, suggesting the closest heart-rhythm follow-up is needed in the first months after discharge.

Review

“Overall, this manuscript reflects a level of methodological rigor that is exemplary for observational research. The authors used a conservative active comparator, prespecified their estimands, decomposed the signal over time, and stress-tested it with competing-risk, landmark, overlap-weighting, E-value, and negative-control analyses. They were disciplined about what the data cannot show, repeatedly declining to claim causality or to endorse routine rhythm surveillance. However, scientific rigor is ultimately judged by whether the central threat to validity is resolved and here a non-null signal in the cataract negative control along with the acknowledged differences in cardiac monitoring between groups leaves surveillance bias as a live alternative explanation for the modest absolute risk difference. This was rigorously conducted and honestly an incredible database study but overall, I would consider this hypothesis generating rather than practice changing with regards to the conclusions. What this work changes is not how we treat endocarditis but how we frame its aftermath. Post-IE rhythm disturbance has been treated as an acute, in-hospital event but this study moves the risk of conduction disease and arrhythmia into more of a longitudinal awareness at least for the first 90 days after diagnosis. Now this data does not change the standard of care by any means but is thought provoking and should encourage more study regarding the first 90 days after diagnosis of infective endocarditis and the impact on conduction and arrhythmias.” - Michael Farbaniec MD FACC, Associate Professor of Medicine Director, Cardiovascular Disease Prevention, Co-Director, Cardiac Rehabilitation Program, Associate Program Director, Cardiology Fellowship, Heart and Vascular Institute, Division of Cardiology, Penn State Health Milton S. Hershey Medical Center. “This is a thoughtfully designed study showing that patients with infective endocarditis remain at increased risk for newly diagnosed arrhythmias and conduction disease after the acute illness, particularly during days 31-90. A particular strength is the use of patients with bacteremia as the comparator group rather than healthy controls. This helps distinguish the risk associated with endocarditis from that of systemic infection more broadly. Although greater cardiac surveillance and underlying heart disease may contribute to the association, the timing is clinically meaningful and warrants heightened attention to rhythm and conduction disorders during early follow-up.” - Emily Guhl, MD, MS, FACC is a board-certified non-invasive cardiologist at Allegheny Health Network Beaver/Sewickley

Medical Researcher Experiences

Based on feedback from the cohort of medical experts, one of the best features of this exploration has been the pace at which AI has accelerated the overall process of their research. In addition to writing the code used to analyze the data in every paper, the experts found that the Muse scaffold was also helpful in synthesizing existing literature, testing whether a hypothesis could be answered with the available data, and even helping produce critiques of collaborators’ work. The experts generated the manuscripts from the first prompt of the model to a manuscript being written in less than two weeks, with some papers taking as few as seven days. Throughout the exploration, participants unanimously felt a significant productivity gain, reporting up to a tenfold improvement in best-case scenarios compared to traditional research processes. We are also sharing a report written by these experts detailing their experience and learnings from doing research with Muse Spark and Muse Wonder.

1 of 7

Looking Ahead

We see this exploration as just the beginning of our work with Muse models in AI for science and health. Although this current study is limited to retrospective analyses, future analyses based on even more comprehensive datasets will allow us to use AI to make inferences and connections among diverse data elements to generate new hypotheses that can be tested in prospective trials. We see AI as playing a pivotal role in advancing scientific research, with models serving as supervised auto-research agents guided by experts making decisions in domains where findings cannot be formally verified. We are excited to continue our exploration in this direction.