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, PhD1,2; W. Frank Peacock, MD, FACEP, FACC, FESC3; Kritika Singh, PhD4; Jared C. Durieux, PhD, MS, MDS, MPH5; Christos Colovos, MD, PhD
- 1Department of Medicine, Division of Infectious Diseases, Penn State Health Milton S. Hershey Medical Center, Hershey, Pennsylvania
- 2Department of Public Health Sciences, Penn State College of Medicine, Hershey, Pennsylvania
- 3Henry JN Taub Department of Emergency Medicine, Baylor College of Medicine, Houston, Texas
- 4Stanford University School of Medicine, Stanford, California
- 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
Return Visits, Admission, and Mortality After Leaving Without Being Seen Among Emergency Department Patients With Behavioral Health Presentations
Sarmed Ashoo, MD1; W. Frank Peacock, MD2; Christos Colovos, MD, PhD
- 1Florida State University College of Medicine, Tallahassee, Florida
- 2Department of Emergency Medicine, Baylor College of Medicine, Houston, Texas
Summary
Ashoo et al. found that among patients with behavioral health emergency department presentations, those who left without being seen had substantially higher rates of return visits (RR 1.86), admission (RR 4.41), and short-term mortality (RR 2.15) than patients who completed care. These findings suggest that behavioral health patients who leave the emergency department without being seen are at significantly elevated risk of acute deterioration and death, identifying a vulnerable population in need of targeted retention and follow-up strategies.
Review
“This extremely thorough retrospective study expands on current research by demonstrating that mental health patients seen in the emergency setting who leave before being seen by a doctor are more likely to have return visits, admissions, and have a higher risk of 30-day mortality. By using Artificial Intelligence to analyze over a million patient cases, a higher impact can be made in both research and clinical spheres. These results point to a continuing need for immediate and innovative interventions in mental health patients in crisis, especially as this at-risk group may already struggle to ask for care. In this study, patients who were high utilizers of the emergency department could count for multiple episodes, so future research should examine whether this pattern continues among people who use the emergency department less frequently.” - Dr. Jonathan Espenan, MD, board-certified psychiatrist at the U.S. Department of Veterans Affairs “This large scale, retrospective study highlights a common problem that affects many emergency departments (ED) across the United States and seeks to better understand what happens to patients who leave the emergency department without receiving care by evaluating the primary outcome of return ED visits within 7 days, and several secondary outcomes including return visits, admissions, and mortality. It focuses specifically on the vulnerable behavioral health population which is helpful to elucidate as the authors highlight, there is a risk of increased morbidity and mortality in this patient population who often lack support and advocacy from family and/or community. The authors were able to show statistically significant results in their primary outcome and also in several secondary outcomes. While not exactly surprising for anyone who works in the field of behavioral health, it does add to the body of evidence that we often lack in our field and hopefully will help healthcare systems better understand the implications of this clinical problem, and start to work on better solutions to prevent some of the underlying issues in the future. This study had several strengths as it had a large patient population and overall strong methodology, however failed to adjust for confounding variables for the main outcomes and/or provide clinically relevant details such as patient diagnoses, severity of illness, or other reasons patients may have left without being seen (i.e. ED acuity, wait times, types of services available, etc); this information could help clinicians better target contributing problems, be able to prioritize the most pressing issues commonly seen, and allocate resources accordingly.” - Dr. Rohini Downs, PharmD, BCPP, board-certified psychiatric pharmacist and Psychiatric Clinical Pharmacy Specialist at Parkland Health & Hospital System in Dallas “The study addresses an important patient safety concern in emergency medicine by comparing outcomes among patients with behavioral health presentations who leave without being seen and those who complete emergency department care. Its large cohort and broad age representation, encompassing both pediatric and adult patients, provide valuable insights into the outcomes of this vulnerable population. The findings have implications for emergency department operations and may inform targeted interventions to improve timely access to emergency care, patient safety, and healthcare delivery. Future research could explore the factors contributing to patients leaving without being seen and evaluate targeted interventions to improve timely access to emergency care and subsequent health outcomes among patients with behavioral health conditions.” - Dr. Enoch Opoku Afriyie, MD, MPH, emergency physician and epidemiologist trained at the Johns Hopkins Bloomberg School of Public Health and Kwame Nkrumah University of Science and Technology
Remission of Type 2 Diabetes in Linked Health Records: Uncommon, Short-Lived, and More Often Ended by Treatment Restart Than by Rising HbA1c
Felipe Lobelo, MD, PhD1; Jared Durieux, PhD, MS, MDS, MPH2; Daniel Kim, MD, PhD3; Christos Colovos, MD, PhD
- 1Emory Global Diabetes Research Center, Rollins School of Public Health, Emory University, Atlanta, Georgia
- 2Evidence Architecture, Stow, Ohio
- 3Division of Cardiology, Department of Medicine, University of Washington, Seattle, Washington
Summary
Lobelo et al. found that documented type 2 diabetes remission was uncommon and usually short-lived in routine care, occurring in 4.3% of previously treated adults and ending more often through treatment restart than through rising HbA1c (47.8% vs. 31.9% within five years), with most restarts occurring while HbA1c remained in the prediabetes range. The findings support reporting glycemic relapse and treatment restart separately and stating how remission and treatment cessation were confirmed, although causation and the reasons for restarting treatment could not be established in this retrospective study.
Review
“This large real-world study provides valuable insights into the frequency and durability of type 2 diabetes remission in routine clinical practice. There are few additional analyses which can make the manuscript more robust. Although the authors report differences in remission rates across racial and ethnic groups, the reasons underlying these variations remain unclear. It would be useful to explore whether these differences persist after adjustment for baseline HbA1c, age, BMI, treatment history and other relevant clinical factors. An additional area worth exploring is whether baseline treatment intensity predicts subsequent diabetes remission. Although the authors report the prevalence of individual glucose-lowering medications, the relationship between the number of medications used at baseline and the likelihood of remission has not been examined. Stratifying remission rates by the number of glucose-lowering drug classes, with adjustment for baseline HbA1c, insulin use and other relevant clinical characteristics, could provide useful insights into whether treatment burden predicts the likelihood of achieving remission in routine clinical practice. It would also be interesting to examine whether the number of glucose-lowering medications used before remission predicts the subsequent risk of relapse. Patients requiring multiple medications may have more advanced diabetes and reduced beta-cell reserve, potentially making sustained remission less likely. An analysis examining the association between prior medication burden and time to glycaemic relapse, separately from treatment restart, would provide valuable insight into predictors of remission durability and help identify patients who may require closer monitoring following treatment withdrawal. One methodological point is especially important: the authors should measure medication burden immediately before treatment withdrawal, rather than relying solely on medications recorded at cohort entry, because treatment may have changed substantially before remission.” - Dr. Harshal Deshmukh, MBBS, MPH, MRCP, PhD, FRACP, endocrinologist and Associate Professor at James Cook University, and Staff Specialist Endocrinologist at Mackay Base Hospital in Queensland. "The importance of this paper lies in how it reframes how the medical community understands type 2 diabetes remission in real-world clinical practice. It shows that, in this cohort, remissions were often cut short not by rising blood sugar, but by proactive clinical decisions to restart medications. Previous real-world studies suggested that type 2 diabetes remission was virtually ‘unsustainable’, reporting high failure rates because they included any end of remission as biological “relapse”. The study underscores that real-world remission estimates are heavily dependent on ‘operational definitions’ (confirmation intervals, handling of unknown treatment gaps, and pharmacy dispensing records). By systematically evaluating these choices, the paper establishes rigorous standards for future observational research. For health systems, researchers, and policymakers designing quality measures or clinical guidelines, the findings emphasize that glycemic relapse and treatment restarts must be reported as ‘separate outcomes.’ Without this distinction, routine-care data cannot be accurately interpreted or fairly compared against intensive clinical trials." - Margarita Ochoa-Maya, MD, board-certified endocrinologist and Medical Director of Rare Disease Endocrinology at Amgen/Horizon Therapeutics.
Pickleball injuries, search interest, and participation in the United States, 2014-2023: an ecological time series study
Daniel Seung Kim1; Felipe Lobelo2; Jared C. Durieux3; Christos Colovos
- 1Division of Cardiology, Department of Medicine, University of Washington, Seattle, Washington
- 2Hubert Department of Global Health, Rollins School of Public Health, Emory University, Atlanta, Georgia
- 3Evidence Architecture, Stow, Ohio
Summary
As pickleball got hugely more popular, Kim et al. found that emergency room visits for pickleball injuries rose sharply from 2014 to 2023, mostly among players in their 60s and 70s. That rise is mainly because far more people are playing, and we found no clear evidence the sport itself has become more dangerous per player. We also showed that Google searches for "pickleball" do not reliably track injuries, so search trends cannot be used as a substitute for real injury counts.
Review
“This is a useful, well-reasoned methods manuscript, and I do believe the two primary discussion points are worth making to the sports medicine community: rising emergency department visit counts aren't the same as rising "per-player" injury risk, and search-interest correlations largely disappear once shared trend and seasonality are removed (the sourdough negative control makes that point memorably). The most interesting point to me in this study are the fact that 70% of injuries occur in 60–79-year-olds, which from a sports medicine perspective, we know that patients in this age group are at a much higher risk of orthopedic injury compared to the young adult population, namely due to predisposing tendinopathy, decreased bone density, decreased dexterity/agility, and medical co-morbidities, however this is likely true across all sports and not specific to pickleball.” - Kaisen Yao, MD, an orthopedic surgeon completing a Sports Medicine fellowship at NYU Langone Health.
Trends in Emergency Department Visits for Electric Scooter and Electric Bicycle Injuries in the United States, 2014–2023: A Retrospective Observational Study
Kritika Singh, PhD1; Jared C. Durieux, PhD, MS, MDS, MPH2; W. Frank Peacock, MD3; James Grichnik, MD, PhD4; Christos Colovos, MD, PhD
- 1Stanford University School of Medicine, Stanford, California
- 2Evidence Architecture, Stow, Ohio
- 3Department of Emergency Medicine, Baylor College of Medicine, Houston, Texas
- 4University of South Florida, Tampa, Florida
Summary
Singh et al. showed the growing strain on emergency care from electric scooters and e-bikes, using national data from US emergency departments.
Review
“I think this is a practical example of how an AI reasoning model can help researchers make sense of large, complex real-world datasets. The clinical takeaway is important too: e-scooter and e-bike injuries are creating a growing burden for emergency departments and health systems, without evidence of a clear increase in injury severity. We all know e-bike use is increasing - but a useful next step would be pairing these injury data with better information about how often people are actually riding e-bikes and e-scooters, so we can determine whether these devices are becoming safer or more dangerous on a per-ride basis.” - Dr. Daniel Seidman, DO, MPH, a board-certified family and sports medicine physician, cofounder of Kin health, and a Clinical Assistant Professor of Family and Community Medicine at Michigan State University College of Osteopathic Medicine. “Overall I think this is a very timely and interesting study, for a topic which is difficult to study because of overlapping or difficult to interpret coding. The study results reinforce my personal impression that e-scooter and e-bike injuries (and associated treatment costs) have increased in recent years. I think this was an interesting use of an LLM to help define the rules for a dataset search that is nonstandard, and then help to validate results, playing to the strengths of such a model. A limitation, however, is that there was no clinician review of the cases, so true accuracy is difficult to assess. I do think there is even further future potential with this line of questioning, as many localities have enacted rules or regulations about micro-mobility devices and a similar approach to data before and after those changes would be very interesting to evaluate whether injury prevention in this space can be legislated (though this would require regional data not currently provided in NEISS).” - Dr. Thomas Falls, MD, an orthopedic surgeon and sports medicine specialist in private practice. “An interesting study investigating the potential impact and burden on emergency and in-patient health services from forms of transport that have gained popularity in recent years. The study highlights an increase in hospital attendances for users of e-scooters and e-bikes in recent years but does not relate this data to an estimated total number of users of both these forms of transport. We are therefore unable to quantify the risk of sustaining an injury requiring hospital attendance for users of e-scooters and e-bikes. Future work could assess this and compare the relative risk to more traditional forms of transport.” - Dr. Oliver Chan, MBBS, MD, FRCS (Tr & Orth), a Consultant Orthopaedic Surgeon specialising in foot and ankle surgery and trauma at Chelsea and Westminster NHS Foundation Trust.
Documented Dental Disease in Patients With Staphylococcus aureus Bacteremia
Paddy Ssentongo, MD, PhD1,2; W. Frank Peacock, MD, FACEP, FACC, FESC3; Christos Colovos, MD, PhD
- 1Department of Medicine, Division of Infectious Diseases, Penn State Health Milton S. Hershey Medical Center, Hershey, Pennsylvania
- 2Department of Public Health Sciences, Penn State College of Medicine, Hershey, Pennsylvania
- 3Henry JN Taub Department of Emergency Medicine, Baylor College of Medicine, Houston, Texas
Summary
Ssentongo et al. showed that dental disease was found about four times more often in patients with potentially fatal Staphylococcus aureus bloodstream (SAB) infections, with the highest rates in patients with prior heart valve infection, opioid use, and injection drug use. The findings support routine mouth checks during SAB care, though causation could not be established in this retrospective study.
Review
“This is a valuable, thoughtfully designed study: medical records noted dental problems far more often in staph bloodstream infections than in E. coli bloodstream infections. The gap appears in every analysis, though closer dental checks during the staph workup may explain part of it, as the authors candidly note. The authors rightly add that the study cannot yet prove the mouth was the source and did not test dental treatment. I think this work puts useful numbers behind the mouth check already in the staph workup, especially when no other source turns up.” - Adam Schaefer,DMD, a CAD/CAM dentist and principal consultant at Strategic Dental Solutions, and former Owner and Clinical Director of 7th and Bell Dental Group and Orthodontics.
Human-Guided Agentic AI for End-to-End EHR Research: Sequential Antibody-Drug Conjugate Pathways in Breast Cancer
Arturo Loaiza-Bonilla, MD, MSEd1,2; Christos Colovos, MD, PhD; Daniel Kim, MD, PhD, MPH3; James M. Grichnik, MD, PhD4; Kritika Singh, PhD5; Paolo Tarantino, MD6
- 1St. Luke's University Health Network, Bethlehem, Pennsylvania
- 2Massive Bio, Inc., Boca Raton, Florida
- 3Division of Cardiology, Department of Medicine, University of Washington, Seattle, Washington
- 4Department of Dermatology and Cutaneous Surgery, University of South Florida, Tampa, Florida
- 5Stanford University School of Medicine, Stanford, California
- 6Dana-Farber Cancer Institute, Boston, Massachusetts
Summary
Loaiza-Bonilla et al. revealed that for the population in our database undergoing treatment for breast cancer, the order of using the antibody-drug conjugates sacituzumab govitecan and trastuzumab deruxtecan lead to equivalent outcomes.
Review
“I found this a genuinely interesting and candid manuscript, particularly because it addresses two areas that are rapidly becoming important in breast cancer research: how we generate meaningful real-world evidence from increasingly large EHR datasets, and how AI agents might support that process from question formulation through analysis and reporting. Its most important message is also reassuring as the ability to automate research increases, clinical expertise becomes more important, not less. The AI can help us interrogate the data, but the clinician still needs to know whether the answer makes clinical sense.” - Dr. Olubukola Ayodele, MBBS, MSc, FRCP, the Medical Oncology Breast Lead at University Hospitals of Leicester NHS Trust and an Honorary Clinical Senior Lecturer at the University of Leicester.
Cladophialophora bantiana: a US clinical case series assisted by artificial intelligence
Ray Y. Chen1; Kaeun Ji1; Christos Colovos1
- 1Meta Platforms, Inc., Menlo Park, California
Summary
In response to a report in the MMWR requesting more information about C. bantiana infections, Chen et al. identified 21 patients with confirmed or probable C. bantiana identified in brain (7) or lung (14) samples and described their clinical characteristics. Among the seven brain infection patients, two died. Artificial intelligence can greatly increase the speed of the analysis of large datasets and make these analyses accessible to more researchers.
Review
"The ability to rapidly detect emerging infectious disease trends is critical to public health surveillance and pandemic preparedness but remains constrained by the time and resources required to analyze large clinical datasets. In response to a recent MMWR report documenting an increase in detection of the rare fungal pathogen, Cladophialophora bantiana, Chen et al. used AI to screen millions of electronic health records for additional cases within just days of the publication. Their analysis did not identify a similar trend in cases which could reflect the difference in the case ascertainment methods used, a major limitation when comparing the two studies. Nonetheless, their findings highlight the speed and feasibility of AI-assisted screening methods for epidemiologic investigations." - Dr. Travis Price, PhD, D(ABMM), a board-certified clinical microbiologist and former Scientific Director at Quest Diagnostics. "To address gaps in knowledge on infections due to Cladophialophora bantiana, the authors utilised AI based technologies to quickly gather data from existing e records of patients with this rare infection. They succeeded in identifying patients and pulling their records for further extracting recorded clinical details, in one day. However, the number of cases is too small to draw conclusions.The search and finalisation of this manuscript were very quick and will be helpful to develop this approach with standard methodologies for defined questions, for future use. This study indicates the need to standardise methodologies for mining specific clinical information from big data sets. It will also be helpful to develop ethical guidelines for accessing such data." - Dr. Elizabeth Mathai, MD, PhD, a clinical microbiologist and global health specialist with more than a decade of technical work at the World Health Organization. "This study demonstrates the potential of AI-assisted analysis to accelerate clinical research by rapidly identifying and characterizing patients with rare and understudied conditions across large electronic health record datasets. The findings provide valuable insights into Cladophialophora bantiana infections, particularly the differences between CNS and pulmonary cases and the potential roles of immunosuppression and underlying health conditions. Future research incorporating larger patient cohorts and independent validation of AI-assisted case identification could strengthen this approach's reliability while enabling the identification of previously unrecognized risk factors, treatment outcomes, and disease patterns." - Dr. Kristina Mahan, PhD, a molecular biologist and synthetic biologist, and former Strain Engineering Scientist at Huue, with postdoctoral research at Sandia National Laboratories and Oak Ridge National Laboratory.
Social Needs Z Codes and Subsequent Acute Care Across 21 U.S. Health Systems
Felipe Lobelo, MD, PhD1; Jared Durieux, PhD, MS, MDS, MPH2; Sarmed Arshoo, MD3; Christos Colovos, MD, PhD
- 1Emory Global Diabetes Research Center, Rollins School of Public Health, Emory University, Atlanta, Georgia
- 2Evidence Architecture, Stow, Ohio
- 3Florida State University College of Medicine, Tallahassee, Florida
Summary
Lobelo et al. found a first Z59 hardship code (a medical-record code documenting housing or financial hardship) predicted more emergency and hospital care in the next year, but administrative-only Z codes also predicted much of it, suggesting the code reflects hardship plus documentation and contact. Coding varied by system after patient-mix adjustment. Risk was highest for homelessness and excess health care burden concentrated in Medicaid. Findings support targeting outreach by specific hardship and reporting hardship risks separately.
Review
“This retrospective study of electronic health and claims records compared three cohorts of adults aged 18 years or older across 21 identifiable U.S. health systems. The first cohort included those with Z59 social-needs codes involving housing stability, income, food access, and transportation. The second cohort had no listed Z59 hardship codes while the third cohort had a Z code unrelated to social hardship. Patients with Z59 hardship codes, particularly those with homelessness codes, as well as Medicaid recipients had the strongest signals for subsequent frequent healthcare utilization. Since coding prevalence ranged from 1.0% to 23.8% in the EHR and claims data, future studies should include examining the factors that led to higher Z59 coding rates to increase sensitivity and specificity.” - Dr. Laura Kahn, MD, MPH, MPP, FACP, physician, One Health consultant, and co-founder of the One Health Initiative, and was previously a Research Scholar in Princeton University’s Program on Science and Global Security. “This study asks whether a Z59 code reflects hardship or simply the act of documentation. The negative-control design using administrative Z02/Z76 codes is a clever approach and the main strength of the paper. The findings are relevant for health systems and payers. However, the admission results are less convincing. In the five systems with mostly valid stay dates, the hardship-specific admission signal was no longer present. The negative-control approach assumes that clinicians assign hardship codes and administrative codes in similar ways. Future studies should test this assumption, possibly with a second control code or chart review, before using the adjustment in practice.” - Dr. Sidath Kapukotuwa, PhD, MS, MCHES, a public health researcher specializing in health disparities and emergency department utilization, and former Program Evaluation Associate at the Kirk Kerkorian School of Medicine at UNLV.
Outcomes in Methicillin-Susceptible and Methicillin-Resistant Staphylococcus aureus Infective Endocarditis Compared With Staphylococcus lugdunensis Infective Endocarditis
Paddy Ssentongo, MD, PhD1,2; W. Frank Peacock, MD, FACEP, FACC, FESC3; Christos Colovos, MD, PhD
- 1Department of Medicine, Division of Infectious Diseases, Penn State Health Milton S. Hershey Medical Center, Hershey, Pennsylvania
- 2Department of Public Health Sciences, Penn State College of Medicine, Hershey, Pennsylvania
- 3Henry JN Taub Department of Emergency Medicine, Baylor College of Medicine, Houston, Texas
Summary
Staphylococcus lugdunensis is often dismissed as a blood culture contaminant but can cause infective endocarditis. Among 4,388 adults, Ssentongo et al. found 30-day mortality was lower with S. lugdunensis (5.4%) than MSSA (7.9%) or MRSA (8.7%), but valve surgery was more frequent (7.7% vs 5.3% and 3.6%). S. aureus caused more embolic and osteoarticular complications. The findings show why S. lugdunensis should never be dismissed as a contaminant and why endocarditis studies should report methicillin susceptibility separately rather than pooling S. aureus.
Review
“This paper addresses a critical gap in literature, specifically outcomes data, regarding infective endocarditis risk status between different organisms and/or susceptibility patterns, notably Staphylococcus lugdunensis. The conclusion made in this article stating that organism and methicillin susceptibility should be reported separately in endocarditis outcome studies is strongly supported by the data reflected in this paper, namely the increased risk of 30-day mortality in the Staphylococcus aureus arm. It should be noted that more patients in the Staphylococcus aureus arm presented with sepsis or septic shock, which could correlate with the increased risk of 30-day mortality. It also appeared that the Staphylococcus aureus arm had more confounders for deep-seated infections, which could increase mortality risk, evidenced by higher rates of osteomyelitis, septic arthritis, and epidural abscesses. For future studies, I would be interested to see a comparison between methicillin-susceptible vs methicillin-resistant Staphylococcus lugdunensis, however the low numbers present in the combined lugdunensis arm could render this analysis hard to power for meaningful results. In regards to data, I would be interested to see information on repeat culture positive rates and durations.” - Mollie VanNatta Day, PharmD, BCPS, BCIDP, clinical pharmacist coordinator and antimicrobial stewardship/infectious diseases clinical pharmacist at Ochsner LSU Health Shreveport in Shreveport, Louisiana. “This study used an LLM to help design and conduct a large analysis of S. lugdunensis, MSSA, and MRSA infective endocarditis. Its most useful finding for clinicians is the higher rate of embolic and osteoarticular complications seen with S. aureus, a reminder to look for metastatic complications. The mortality claims are overstated; the supplement itself states the analysis “does not establish a mortality difference,” and the paper does not address existing literature which reports S. lugdunensis endocarditis mortality of roughly 20–42%. Future studies would benefit from Duke-confirmed endocarditis, culture-confirmed organism identification, endocarditis-related severity measures such as the presence and size of vegetations, and careful human review of any LLM-assisted analysis.” - Storm Dorrough, MD, board-certified pediatrician and pediatric infectious disease physician at the University of New Mexico Children’s Hospital in Albuquerque. “The authors used a large EHR database with AI assistance to study a substantially larger patient cohort with S. lugdunensis endocarditis than earlier reports, which is a useful contribution for an uncommon infection. I also appreciated the separate comparisons with MSSA and MRSA, although I am less convinced by the mortality conclusions because the results change in the sensitivity analyses. The main limitation is that the groups were identified using different diagnostic criteria, making it difficult to tell whether the outcome differences reflect the infections themselves or how patients were selected. Future studies can define these populations well to establish in greater details the risk of this high mortality bug.” - Ali Ahmad, MD, board-certified internist and cardiovascular diseases fellow at Baylor College of Medicine, and a longtime research collaborator in cardiovascular medicine at Mayo Clinic.
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.
“What excites me most about this project is that it concentrates the best parts of science - the curiosity, creativity, collaboration, and bouncing ideas around - while removing so much of the friction.
Our team created an ideation skill in Muse that let me explore the many ideas in my head with little opportunity cost. I was able to quickly evaluate which questions were worth pursuing and save myself the dread of realizing the idea you just sunk weeks or months into is a dead-end.
It was also really helpful that we could review each other's sessions you could get lots of eyes on a problem quickly, and Muse could learn from what others uncovered. It's a new way to collaborate, but incredibly useful.”
“With the research environment we're using, the biggest change for me is how freely I can explore an idea. Getting into the EMR data is almost effortless, and when a question reaches outside radiology, it pulls the relevant guidelines and literature and fills in the gaps. I can go from a question to a real sense of whether it's worth pursuing in an hour or two, which frees me up to spend more time on the ideas themselves.
What I appreciate about Muse Wonder is that it isn't sycophantic. Some other models I've used will make an idea sound great in the moment, and only later do I realize it didn't hold up. Muse Wonder is willing to call me out and steer me toward a better approach, often before I've spotted the problem myself. That saves a lot of time, and it's a big part of why I trust it.”
“The main lesson for me from working on this project is that AI makes the research faster and more accessible. Faster because AI can collate and process data much faster than we can. Accessible because now any programming-challenged researcher like me can chat with the agent and do the research by myself from my desk. I don't need to wait for the data engineers and statisticians to pull and analyze the data for me, as they are all busy too. What's novel here is not the research topics being analyzed but the research process to conduct the analyses. The research process is what will be fundamentally changed by AI going forward.”
“AI can help us develop and refine research ideas, explore questions in the data, and consider different interpretations of the findings, with researchers guiding and critically evaluating the work at every step. For people balancing research with training, clinical work, and other responsibilities, that support can make it easier to stay involved. To me, that is what it means for AI to help democratize research: lowering practical barriers without replacing experts or rigorous review.
I saw that with Muse Wonder. One thing I really appreciated was that it understood a paper is never just the main manuscript. It also includes the supplement, tables, figures, reporting checklists, and multiple rounds of revision. When the analysis changes, all of those pieces need to change with it. Muse Wonder helped me keep the entire package organized and consistent as the work evolved. I could spend less time reorganizing documents and more time thinking carefully about the science. I still read every line, checked every scientific claim, and made the final decisions, but the path from an initial idea to a submission-ready manuscript felt far more manageable.”
“As a practicing oncologist, I see important research questions every day that we lack the time and resources to investigate. The Cambrian explosion of medical knowledge, along with ever changing data needed to be applied to our daily practice and future research are making it paramount to have Superintelligence at our hands to support our workflows. Leveraging Muse Wonder really helped us work within a de-identified database of nearly five million patients, connecting cohort construction, executable analyses, data-quality checks, literature review, and manuscript development in one physician-directed, agentic workflow. We could challenge assumptions, refine interpretations, and carry those corrections through the analysis and the paper.
Moving from a clinical question to a manuscript in days opens the door to pursuing questions we might otherwise leave unanswered. We remain accountable for the science, while our capacity to contribute expands dramatically. We are beginning to close the gap between the questions we bring back from the bedside and the research we can actually accomplish.”
“I have used Muse Wonder as a research partner to work on research questions in large-scale clinical, claims, environmental and population-health datasets. It helped refine hypotheses, design cohorts and target trial-simulation analyses, identify potential sources of bias, write and troubleshoot analytic code, and draft manuscripts.
I still provide the clinical and epidemiological judgment, challenge assumptions, refine definitions and determine which analyses are scientifically meaningful, but the model has allowed me to test more ideas, iterate faster and pursue questions that would otherwise require substantially more time and technical support. Its greatest value has been shortening the distance between a promising idea and a rigorous scientific product.”
“What has changed most for me isn't necessarily the research itself, but how we conduct it. Traditionally, a considerable amount of time is spent translating research questions into analytic specifications, writing and debugging code, validating outputs, and going back and forth between methodological decisions and implementation. With AI, much of that process can happen far more efficiently and iteratively, allowing me to focus more of my attention on the scientific questions, study design, causal assumptions, potential sources of bias, and interpretation of findings.
What I find particularly exciting is that AI is beginning to close the gap between scientific reasoning and technical execution. I can explore methodological alternatives, interrogate assumptions, identify potential problems, and refine an analytic approach without every iteration requiring a separate programming cycle. That doesn't eliminate the need for epidemiologists, statisticians, clinicians, or independent validation. If anything, it makes our expertise even more important because the speed at which analyses can be conducted doesn't necessarily translate into the credibility of the evidence being generated.
For me, the real advancement is not simply that AI allows us to do research faster, but that it creates opportunities to conduct research in a more iterative, comprehensive, and scientifically rigorous way than traditional workflows have generally allowed.”
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.