When the Diagnosis Arrives by App: Why Most Patients Still Want a Human Voice
Key Takeaways:
- In a UT Southwestern survey of more than 2,400 people diagnosed with cancer, 75% said they would prefer to receive the news directly from their physician – in person or by telemedicine – rather than through an electronic patient portal.
- More than half of those who first learned of their diagnosis via the portal were alone at the time, often without support from a clinician or family member.
- Researchers are calling for a more personalised approach, including better portal notification settings, tiered or delayed release of sensitive findings, and plain-language summaries of radiology and pathology reports.
A digital convenience with an emotional cost
Electronic patient portals have transformed how quickly people can see their own test results. For routine bloodwork or a clear scan, near-instant access is a welcome convenience. But the same speed raises a difficult question for clinicians: how should a new cancer diagnosis be communicated when a patient can open the result on a phone before anyone has had the chance to talk it through?
That tension has grown sharper since 2021, when a provision of the 21st Century Cures Act came into force in the United States. The regulation requires that patients have timely, unrestricted access to their electronic health information – which, in practice, means a growing number of people are discovering a new or recurrent cancer diagnosis through their portal, sometimes with no clinician present to interpret it or answer the questions that immediately follow.
What the UT Southwestern survey found
A new survey carried out at UT Southwestern Medical Center suggests that, for most people facing a cancer diagnosis, faster is not better. The findings, published in JAMA Network Open, show that 75% of respondents would prefer to learn about a cancer diagnosis directly from their physician, whether in person or through a telemedicine appointment.
The 2025 survey gathered responses from more than 2,400 people who were diagnosed with cancer at the Harold C. Simmons Comprehensive Cancer Center between 2019 and 2023, giving the researchers a substantial real-world picture of how patients want sensitive results delivered.
According to the study’s lead author, Sheena Bhalla, M.D., Assistant Professor of Internal Medicine in the Division of Hematology and Oncology and a medical oncologist at the Simmons Cancer Center, the broad enthusiasm for digital access does not extend neatly to oncology. “While most patients in the general population appreciate rapid electronic access to test results, the situation for patients with cancer is much more nuanced,” she said. “Learning about a cancer diagnosis without the ability to immediately ask questions or discuss next steps with a trusted clinician can add to the significant stress, uncertainty, and fear that patients experience.”
Preferences are not one-size-fits-all
The survey also revealed that there is no single right way to share a result. Preferences varied according to people’s prior experiences, how frequently they used their portal, and their demographic characteristics. Men, for instance, were more likely than women to prefer learning of a diagnosis through the portal.
For senior author David Gerber, M.D., Professor of Internal Medicine in the Division of Hematology and Oncology and of Epidemiology in the Peter O’Donnell Jr. School of Public Health, and co-Director of the Simmons Cancer Center Office of Education and Training, that variation is precisely the point. “These findings highlight the need for a more personalized, tailored approach to communicating sensitive and life-changing results,” he said. “Moving beyond a one-size-fits-all approach can help clinicians provide a more thoughtful, compassionate patient experience.”
The hidden consequence: facing the news alone
Perhaps the most striking insight concerns the circumstances in which people are receiving these results. Among those who learned of their diagnosis through the portal, more than half reported that they were alone when they read it.
Dr Bhalla described this as one of the most troubling side effects of real-time access. “That’s one of the most unintended consequences of real-time access,” she said. “Patients are often alone without support from their physician or family at one of their most vulnerable moments.”
Possible solutions for clinicians and health systems
The researchers are clear that the answer is not to roll back access, but to design around it more thoughtfully. They point to several potential measures, including raising awareness among both clinicians and patients of the portal notification settings already available; developing tiered or delayed-release approaches for particularly sensitive findings; and integrating supportive digital tools such as plain-language summaries for radiology and pathology reports.
Policy is beginning to catch up. Since the Cures Act took effect, three states – including Texas – have enacted laws permitting the delayed portal release of cancer-related and other sensitive results, giving care teams a window to reach out before a patient is left to interpret difficult news on their own.
Looking ahead
For the study’s authors, the work is a starting point rather than a conclusion. “Further study and increased interdisciplinary collaboration among oncology clinicians, health services researchers, and digital health experts can help us better understand how patients receive and react to cancer diagnoses,” Dr Bhalla said. “Our goal is to increase awareness of this issue and help drive innovative approaches to patient-centered communication.”
Source: UT Southwestern Medical Center
Read More
Digital Health Tools Are Now a Routine Part of Everyday Care in the US
Key Takeaways:
- A landmark review of more than 8 billion interactions across US healthcare finds that online portal messaging has become a standard, everyday part of care rather than an occasional add-on.
- Digital communication is supplementing in-person medicine, not replacing it – office visits have rebounded to two to three per patient each year while portal messages have more than doubled.
- Researchers warn that the growing digital workload sits on top of clinicians’ existing duties, raising new questions about staffing, training and the role of AI support tools.
A new picture of how Americans reach their clinicians
At least 12 per cent of people in the United States now contact their healthcare providers about appointments, test results and ongoing treatments through secure online patient portals and health apps, according to a major new study. At the same time, traditional in-person visits to the doctor’s office have recovered from their pandemic-era decline. The findings suggest that while digital medicine has become a routine feature of care, it is adding to in-person services rather than displacing them – an evolution that researchers say is reshaping how hospitals and clinics run day to day.
These are the central conclusions of a study led by researchers at NYU Langone Health, described as the largest review ever conducted of communications recorded in Epic electronic health records. The team analysed more than 140 million patient records drawn from 2,067 hospitals and 47,100 health clinics across the US, examining over 8 billion interactions between patients and providers that took place between January 2020 and December 2025.
What the data showed
Published online in the Journal of the American Medical Association (JAMA) on 22 June, the study found that online portal messages more than doubled between 2020 and 2025, rising by 153 per cent. Over the same period, total telephone calls fell by 6 per cent. The number of people with an active Epic health record climbed from 94 million in 2020 to 140 million in 2025. During the first three months of 2025, 30 per cent of active Epic patients – some 42 million people – sent a portal or health app message to their clinician.
Crucially, this surge in portal activity is not coming at the expense of face-to-face care. In-office visits have returned to an average of between two and three per patient each year. Messages from patients to their providers have, meanwhile, doubled since the pandemic, increasing from an average of 2.2 per year in early 2020 to 5.4 per year in late 2025.
“Our study shows that use of patient portals, health apps, and messaging are now a routine part of everyday patient care across America, not simply side channels used occasionally,” said study senior investigator Michal A. Mankowski, PhD.
Dr Mankowski, an assistant professor in the Department of Surgery at NYU Grossman School of Medicine, said the findings show that people now have far more direct access to physicians and other clinicians than before.
“Our findings reveal that while digital health tools have become a core part of healthcare, delivery is becoming more continuous and timeless, and no longer tied to scheduled appointments during routine work hours,” said Dr Mankowski.
The scale of digital care since 2020
The review also quantified the sheer volume of activity logged through Epic record systems since 2020. Over that period, people in the US booked at least 1.77 billion in-person visits to health clinics, sent 1.34 billion messages to their providers and received roughly 3.25 billion portal messages from providers in return. Epic systems also documented 1.59 billion telephone calls and 146 million virtual telehealth portal visits.
A new layer on top of clinical work
Study co-investigator Dorry L. Segev, MD, PhD, said the digital delivery of healthcare does not replace established ways of working; rather, it adds a further layer of steps to existing workflows. To cope with this new reality, he argued, hospitals, clinics and healthcare workers will need to plan ahead for staffing and support.
“Modern delivery of healthcare means increasingly that healthcare providers will have to balance their digital workload on top of their traditional clinical workload,” said Dr Segev, a professor and vice chair in the Department of Surgery at NYU Grossman School of Medicine.
“Clinical staff will need to be trained in mastering the tools of messaging in healthcare; in using AI support programs, including chatbots that can frame content to minimize its complexity; and in making the most effective use of clinician time needed for online billing and online counseling,” added Dr Segev, who is also a professor in NYU Grossman’s Department of Population Health.
He noted that NYU Langone already uses AI support tools to speed up the drafting of physician and provider notes. Looking ahead, Dr Segev said the team plans to examine digital-use trends within individual healthcare systems, including NYU Langone, in order to identify regional and outpatient clinic-specific shifts that could affect operational planning.
How the study was carried out
For the research, the team drew on Epic Cosmos, a national dataset containing the electronic health records of more than 300 million patients in the US. The dataset includes information from a majority of the hospitals and clinics that use Epic, the country’s largest vendor of electronic health record systems. Epic had no role in carrying out the study. Funding was provided by NYU Langone.
Alongside Dr Mankowski and Dr Segev, the NYU Langone researchers involved were lead investigator Jane J. Long, MD, and co-investigators Mara A. McAdams DeMarco, PhD; Mark D. Schwartz, MD; Joshua Chodosh, MD; and Eric K. Oermann, MD.
Disclosures
Dr Mankowski was recently elected to serve on the governing board of Epic Cosmos. Dr Schwartz reported being president-elect of the Society of General Internal Medicine. Dr Segev has received consulting and/or speaking honoraria from Sanofi, CareDx, Moderna, AstraZeneca, Roche, Optum, OrganOx, Hansa and Biosidus, and is a journal editor for Springer. None of these activities are related to the current JAMA study. NYU Langone is managing the terms and conditions of these relationships in accordance with its policies and procedures.
Source: NYU Langone
Read More
Autonomous AI Agent Matches and Exceeds Physicians Across Simulated Electronic Health Record Cases
Key Takeaways:
- MIRA is an autonomous AI agent that diagnoses and plans treatment inside a simulated electronic health record, rather than acting as a narrow chat tool.
- It reached 88.9% diagnostic accuracy across 574 cases, outperforming board-certified physicians (78.1%) and a mixed-seniority team (71.1%).
- Safety results were strong but preliminary, and the authors stress that MIRA is not a replacement for human clinicians.
A new kind of medical AI agent
A recent study published in the journal Nature introduced MIRA, an autonomous AI agent designed to operate within sandboxed EHR environments. Rather than acting as a single-purpose assistant, MIRA uses a suite of digital tools to simulate the full arc of a clinical workflow. It can order tests, synthesise the results, and produce diagnoses and treatment plans, all while communicating through a chat interface with a patient AI agent that is grounded in the documented history of present illness extracted from retrospective notes from genuine cases.
The system runs on a Fast Healthcare Interoperability Resources (FHIR) based architecture, which executes the agent’s tool calls and records its medical outputs. The researchers note that the example data presented in the paper were shortened and slightly modified to comply with the privacy restrictions attached to the dataset.
Unlike earlier implementations, which were predominantly task-specific chat applications, MIRA was built to independently take in patient histories, order the relevant diagnostic tests, and then use those datasets to reach diagnoses and treatment plans within a controlled simulation. Across the 574 MIMIC-IV cases, MIRA achieved 88.9% diagnostic accuracy, and in a matched 311-case physician comparison it reached 87.8% accuracy, significantly outperforming experienced human physicians under identical simulated conditions while demonstrating strong, though not perfect, safety and guideline performance.
Background: from passing exams to working a ward
Large language models (LLMs) have already proven highly capable at passing standardised medical examinations and answering complex clinical questions. Reviews of the field show, however, that translating this raw clinical knowledge into the operational workflow of a hospital has remained a major challenge.
This gap is attributed to the architectural design of traditional medical AI tools, which behave as narrow, task-specific search or text-generation utilities rather than as active partners in care. By contrast, true clinical decision-making is characterised as an intricate, multi-step process in which doctors repeatedly interview the people in their care, order blood tests or imaging, synthesise conflicting results, and update their hypotheses before arriving at a final treatment plan.
Nearly all of this clinical work takes place within EHR systems that rely on complex, standardised coding protocols. Until now, it remained unproven whether an automated system could reliably handle this end-to-end clinical action space in a realistic, EHR-style environment without committing unacceptable errors.
About the study
The study set out to address this functional gap by developing MIRA, a novel AI tool designed to autonomously ingest and access medical records, identify knowledge gaps, and order diagnostic tests to supplement the EHR record, before using the completed dataset to recommend clinical interventions.
The researchers then tested MIRA’s capabilities in a sandboxed, virtual EHR environment compliant with standard healthcare protocols, including HL7 FHIR. The sandboxed test was conducted on a curated benchmarking dataset of 574 real-world emergency department cases from the Medical Information Mart for Intensive Care (MIMIC-IV) database.
The cases included spanned eight distinct diagnoses across surgery (appendicitis), internal medicine (pneumonia), and oncology (pancreatic cancer), which MIRA navigated using 11 specialised digital tools offering more than 85,000 operational choices. The agent was permitted to request physical examinations, order targeted laboratory values, look up medical histories, and generate medication orders within the simulated EHR, rather than in live patient care.
How MIRA was compared with clinicians
MIRA’s output was compared against two distinct groups of human physicians managing exactly the same cases under identical conditions. The first group was a cohort of four board-certified physicians. The second was a mixed-seniority team consisting of four residents and two board-certified doctors.
A separate, conventional text-based AI agent was used to simulate the people under MIRA’s care, and under the care of the human physician teams. This agent was instructed to respond to questions posed by MIRA or its human counterparts solely on the basis of authentic clinical histories, while resisting adversarial attempts to trick it into prematurely leaking information. The authors noted, however, that simulated patient speech may be more structured than real emergency department conversations.
Study findings
The results revealed that MIRA performed at or above the level of experienced human doctors. It achieved 88.9% diagnostic accuracy across the full 574-case dataset and 87.8% accuracy in the matched 311-case physician comparison. By comparison, the board-certified physicians reached an average accuracy of 78.1% (p < 0.001), while the mixed-seniority medical cohort averaged 71.1% (p < 0.001).
MIRA was found to excel at identifying appendicitis and pancreatitis, achieving a perfect 100% recall for laparoscopic appendectomies. For pancreatic cancer, its diagnostic performance was equivalent to that of the board-certified physicians, while pneumonia and urinary tract infections remained more challenging.
Accuracy without simply “ordering everything”
Notably, MIRA did not achieve its superior accuracy by simply “ordering everything”. While it was observed to request a broader, more comprehensive set of individual blood parameters than the human doctors, its overall test selection remained well below the historical baselines recorded in the dataset.
The findings further demonstrated that the model successfully avoided the systematic over-ordering of high-cost radiological imaging, matching or exceeding physicians on overall resource-alignment metrics.
Safety performance
The safety evaluations were similarly encouraging, though still preliminary. An independent, blinded medical review of 56 patient-level outputs, together with a separate assessment of 468 prescriptions written by MIRA, established that the agent caused zero high-severity drug–drug interactions, zero renal dosing incompatibilities, and zero medication-allergy mismatches. Route specification was the weakest prescription field, at 97% correctness.
When making critical hospital admission decisions for pneumonia and pulmonary embolism, MIRA achieved a perfect recall score of 1.00, indicating that it never missed a single person who required inpatient care. The pulmonary embolism analysis did, however, suggest a tendency towards over-admission, reflecting a cautious disposition strategy.
Conclusions
The study introduces an integrated EHR AI agent, MIRA, that successfully translates clinical intents into structured, safe, and accurate operations, with the potential to support physicians in their work. The authors are careful to caution, however, that MIRA and similar AI agents are not replacements for expert human staff.
The model did not reach 100% perfection across all treatment choices, such as specific antibiotic selections, which highlights the ongoing need for strict human supervision and patient-level safeguards. Future iterations of the model may improve their performance by incorporating evidence from retrieval-based support, stronger governance, and prospective real-world validation before any clinical deployment.
Read More
AI Models Identify Hidden Cardiac Arrest Risk in Routine Patient Data
Key Takeaways:
- Researchers have built AI models that analyse electronic health records and electrocardiograms to identify people at elevated risk of sudden cardiac arrest, which kills more than 400,000 Americans each year.
- In a real-world group of nearly 40,000 patients, the combined model correctly flagged 153 of 228 high-risk people who later experienced cardiac arrest, narrowing risk prediction from 1 in 1,000 to 1 in 100.
- The models also surfaced modifiable contributors such as electrolyte disorders, substance use and medication interactions, pointing to practical opportunities for clinicians to intervene.
A new approach to an unpredictable emergency
Researchers have developed artificial intelligence (AI) models capable of analysing electronic health records (EHR) and electrocardiograms to pinpoint people in the general population who face a heightened risk of sudden cardiac arrest. The condition is responsible for more than 400,000 deaths each year in the United States and carries a survival rate of just 10%, making any tool capable of forecasting it a meaningful step forward.
The work represents a notable advance in anticipating an event that is widely considered difficult, if not impossible, to predict, and which often strikes people who have no previously known heart disease.
“Using artificial intelligence applications and health records data, the prediction of cardiac arrest in the general population is feasible,” said Dr Neal Chatterjee, the study’s lead investigator and a cardiologist at the University of Washington School of Medicine.
The paper was published on 11 May in JACC: Advances, a journal of the American College of Cardiology. Additional co-senior authors are affiliated with Massachusetts General Hospital and the Broad Institute of MIT and Harvard.
How the models were built
The investigation drew on a test population of roughly 1.7 million patients enrolled in a large healthcare system in the United States. The team built three separate AI models, each trained on a distinct dataset. The first, referred to as “EKG-only,” relied solely on electrocardiogram readings. The second, “EHR-only,” weighed 156 clinical features drawn from patients’ health records. The third combined both EKG and EHR data into a single integrated model.
The researchers developed and validated their models across three distinct patient groups.
Training cohort
The models were initially trained using data from 993 people who had experienced out-of-hospital cardiac arrest between 2013 and 2021, alongside 5,479 control patients matched for age and sex who had not. This stage allowed the AI to learn which patterns in EHR entries and EKG readings were linked to a higher risk of cardiac arrest.
Testing cohort
To confirm that the models could reliably distinguish between high- and low-risk indicators, the researchers applied them to a separate group consisting of 463 cardiac arrest cases from 2022 to 2023 and 2,979 control patients. The risk associations identified in this testing group closely mirrored those established during training.
Real-world cohort
The final stage involved 39,911 people who had received EKGs during 2021, regardless of their health status. The researchers examined the records of those within this group who went on to experience cardiac arrest over the following two years, assessing how closely their profiles aligned with the risk patterns identified by the models.
Within this real-world group, the combined EHR-EKG model accurately predicted 153 of 228 people who were classified as high-risk and who later went on to experience a cardiac arrest.
Bringing theoretical risk into focus
The shift in predictive precision is one of the study’s most striking outcomes.
“With these models, we’re able to enrich risk prediction from about 1 in 1,000 down to 1 in 100,” Chatterjee said. “If your doctor were to tell you that your risk of cardiac arrest is 1 in 100, that would catch your attention. We’re bringing a theoretical risk into focus.”
Another encouraging finding concerned the performance of the EKG-based model on its own. AI-enhanced analysis of electrocardiograms alone demonstrated strong predictive ability, only modestly behind the two models that drew on EHR data.
“The 12-lead EKG is a low-cost tool that might stratify patients’ risk for cardiac arrest in any community around the world,” Chatterjee said.
Risk factors beyond traditional cardiology
The study also surfaced risk factors that lie outside the conventional cardiovascular picture. Contributors flagged by the models included electrolyte disorders, substance use and interactions between medications, all of which are often addressable through clinical attention.
“We show some relatively low hanging fruit … modifiable risk factors,” Chatterjee noted. “A model that flags a patient as high-risk might prompt somebody taking care of a patient to review their medical history and their medications.”
Open questions for clinical practice
While the results demonstrate that predicting cardiac arrest risk at the population level is achievable, Chatterjee was careful to note that the next stage of inquiry involves working out what clinicians should actually do once a patient is flagged.
“We need to figure out which follow-on studies to pursue to understand what we do with this patient information. What screening, what surveillance, what intervention is warranted?”
Limitations of the study
Several constraints temper the findings. All of the data was drawn from a single healthcare system, leaving open the question of whether the models would perform similarly across populations with different demographic profiles or patterns of care. The real-world group was also restricted to people who had received an EKG, and these individuals may differ in important ways from those who had not undergone such testing. In addition, the AI-enhanced interpretations of EKGs could reflect biases tied to demographics or to the way care is delivered.
Funding and support
The research received support from the National Institutes of Health (K23HL169839, R01 HL160003, R01 HL168889, K24 HL153669, R01HL092577, R01HL157635), the American Heart Association (23CDA1050571, 961045), the European Union (MAESTRIA 965286) and the Foundation Leducq (24CVD01). Chatterjee is supported through a philanthropic donation from Kevin and Ann Harrang and through the John and Cookie Laughlin Endowed Professorship.
Source: UW Medicine
AI Analysis of Health Records May Help Identify ADHD Risk Years Earlier
Key Takeaways:
- Artificial intelligence can analyse routine electronic health records to estimate a child’s risk of developing ADHD years before diagnosis.
- The model demonstrated strong accuracy across diverse populations, using data from more than 140,000 children.
- The tool is designed to support earlier evaluation and intervention, not to replace clinical diagnosis.
AI and the challenge of delayed ADHD diagnosis
Attention-deficit/hyperactivity disorder (ADHD) affects millions of children worldwide. Despite its prevalence, many children experience significant delays before receiving a formal diagnosis. This can limit access to timely support, even when early signs are present.
New research from Duke Health suggests that artificial intelligence may help address this gap. By analysing routinely collected electronic health records, researchers have developed a tool capable of identifying patterns that indicate a higher likelihood of future ADHD diagnosis, potentially years in advance.
Unlocking insights from routine healthcare data
The study, published in Nature Mental Health on April 27, demonstrates how existing healthcare data can be used to support earlier clinical decision-making.
“We have this incredibly rich source of information sitting in electronic health records,” said Elliot Hill, lead author of the study and data scientist in the Department of Biostatistics & Bioinformatics at Duke University School of Medicine. “The idea was to see whether patterns hidden in that data could help us predict which children might later be diagnosed with ADHD, well before that diagnosis usually happens.”
Rather than relying on new or specialised testing, the approach draws on information already collected during standard healthcare visits. This includes developmental milestones, behavioural observations, and clinical events recorded from birth through early childhood.
How the AI model was developed
To build and test the model, researchers analysed electronic health records from more than 140,000 children, including both those diagnosed with ADHD and those without the condition.
The AI system was trained to identify combinations of factors that tend to appear before an ADHD diagnosis is made. Over time, it learned to detect subtle patterns across large datasets that may not be easily recognised through conventional clinical assessment alone.
The model showed strong performance in estimating ADHD risk in children aged 5 years and older. Notably, its accuracy remained consistent across different population groups, including variations in sex, race, ethnicity, and insurance status.
A support tool, not a diagnostic replacement
The researchers emphasise that the AI system is not intended to diagnose ADHD. Instead, it serves as a clinical support tool that can help identify children who may benefit from closer monitoring or earlier referral for specialist assessment.
“This is not an AI doctor,” said Matthew Engelhard, M.D., Ph.D., in Duke’s Department of Biostatistics & Bioinformatics, and senior author of the study. “It’s a tool to help clinicians focus their time and resources, so kids who need help don’t fall through the cracks or wait years for answers.”
By highlighting children who may be at higher risk, the tool could enable healthcare professionals to prioritise evaluation and initiate discussions with families sooner.
Potential benefits of earlier identification
Earlier identification of children at risk of ADHD could have meaningful implications for long-term outcomes. Research consistently shows that timely diagnosis and intervention are associated with improved academic performance, social development, and overall health.
“Children with ADHD can really struggle when their needs aren’t understood and adequate supports are not in place,” said study author Naomi Davis, Ph.D., associate professor in the Department of Psychiatry and Behavioral Sciences. “Connecting families with timely, evidence-based interventions is essential for helping them achieve their goals and laying a foundation for future success.”
The ability to flag potential concerns earlier may also help reduce the delays that many families face when seeking answers and support.
Next steps and ongoing research
While the findings are promising, the researchers stress that further validation is needed before such tools are implemented in routine clinical practice. Additional studies will be required to confirm effectiveness, assess real-world impact, and ensure safe integration into healthcare systems.
Hill and Engelhard have also explored the broader use of AI models in identifying risks and contributing factors for mental health conditions in adolescents, signalling a growing interest in predictive tools within this field.
Study authors and funding
In addition to Elliot Hill, Matthew Engelhard, and Naomi Davis, the study authors include De Rong Loh, Benjamin A. Goldstein, and Geraldine Dawson.
The research was supported by grants from the National Institute of Mental Health (K01-MH127309, UL1 TR002553) and the National Center for Advancing Translational Sciences.
Source: EurekAlert!
Read More
Large Language Models Show Promise in Detecting Drug Safety Signals from Clinical Notes
Key Takeaways:
- Large language models can identify immune-related adverse events in clinical notes without task-specific training, offering a potential alternative to labour-intensive manual review
- Performance remains below the threshold required for clinical decision support, with models tending to overpredict adverse events
- Despite limitations, this approach may support large-scale safety monitoring and accelerate research into cancer immunotherapies
The challenge of detecting drug safety signals
Drug safety signals are often embedded within unstructured clinical text, particularly in electronic health records. Identifying these signals has traditionally required either manual chart abstraction, which is resource-intensive, or natural language processing systems tailored to specific drugs and healthcare settings.
This challenge is particularly evident in the case of immune checkpoint inhibitors. These cancer therapies, first introduced in 2011, are associated with a broad range of immune-related adverse events. These events can affect multiple organ systems, including the colon, liver, lungs, heart, nervous system, skin, and endocrine system, making systematic detection complex and time-consuming.
Exploring large language models as a solution
Large language models are increasingly being explored as a way to streamline the identification of drug safety signals within clinical text. A multicentre study, published in eBioMedicine, evaluated whether these models could detect immune-related adverse events associated with immune checkpoint inhibitors.
The study focused on a zero-shot learning approach. In this setting, the model receives a single, detailed prompt without prior examples. The prompt used by the researchers began: “You are a clinical expert in identifying immune-related adverse events caused by immune checkpoint inhibitors …” and included a list of six immune checkpoint inhibitors alongside numerous associated adverse events.
This prompt was applied to clinical notes from multiple sources. These included records from 100 people treated at Vanderbilt Health, 70 people from the University of California, San Francisco, and 272 people enrolled in seven Roche-sponsored clinical trials.
Study design and model performance
The research team evaluated three models: GPT-3.5, GPT-4, and GPT-4o, with GPT-4o demonstrating the strongest overall performance.
To assess accuracy, the investigators used F1 scores, a metric that balances false positives and false negatives. Scores range from zero to one, with values above 90 percent considered excellent. A score of 80 percent or higher may be sufficient for use in automated clinical decision support systems.
At the patient level, GPT-4o achieved average F1 scores of 56 percent for Vanderbilt Health data, 66 percent for University of California, San Francisco data, and 62 percent for Roche clinical trial data. The models showed a consistent tendency to overpredict the presence of immune-related adverse events.
When analysing individual clinical notes, the model achieved an average F1 score of 57 percent across 667 notes from Vanderbilt Health, evaluating 17 different adverse events.
Implications for clinical practice and research
The findings suggest that large language models can play a role in identifying drug safety signals, even without task-specific training data.
“Manual patient chart abstraction for monitoring the safety and efficacy of drugs already at market requires tremendous resources and puts a drag on the pace of discovery in precision medicine. And that’s especially true with immune checkpoint inhibitors, where the adverse events are so varied. If zero-shot learning with LLMs could help with these notes, it could significantly reduce time and costs for all concerned,” said the report’s corresponding author, Cosmin Bejan, PhD, assistant professor of Biomedical Informatics at Vanderbilt Health.
However, the current level of performance falls short of what would be required for clinical decision support.
“These results show that zero-shot learning with a powerful LLM is useful for detecting these adverse events,” Bejan said. “This performance does not rise to the level required for clinical decision support, but the method could be valuable for automated irAE extraction across multiple sites, potentially speeding discovery and enhancing the safety and effectiveness of cancer immunotherapies.”
Wider research context
The study involved collaboration among multiple researchers at Vanderbilt Health, including Yaomin Xu, PhD, Eric Mukherjee, MD, PhD, Matthew Krantz, MD, Douglas Johnson, MD, MSCI, Elizabeth Phillips, MD, and Justin Balko, PhD. Funding support was provided in part by the National Institutes of Health.
Related research further highlights safety concerns associated with immune checkpoint inhibitors. In a research letter published in JAMA Oncology, Mukherjee, Phillips, and colleagues used logistic regression analysis of adverse event reports from the Food and Drug Administration. They confirmed that these therapies are independently associated with an increased risk of Stevens-Johnson syndrome and toxic epidermal necrolysis, which are severe and potentially life-threatening skin reactions. The study also found that this risk may be linked to exposure to human leukocyte antigen–restricted drugs.
Conclusion
Large language models represent a promising tool for extracting clinically meaningful insights from unstructured health data. While their current performance limits direct clinical application, their ability to operate across multiple datasets without task-specific training suggests potential for supporting large-scale pharmacovigilance efforts. As these models continue to improve, they may contribute to more efficient and comprehensive monitoring of drug safety in clinical practice.
Read More
Growing Use of Emojis in Electronic Health Records Raises Safety Questions
Key Takeaways:
- Emoji use in electronic health records has increased steadily between 2020 and 2025, appearing in thousands of clinical notes.
- While emojis may offer a quicker, more expressive way to communicate, they introduce risks of misinterpretation across clinicians and patients.
- Experts suggest that clearer governance and guidance may be needed to ensure safe and consistent use in clinical documentation.
Study reveals growing presence of emojis in clinical records
A recent study from Michigan Medicine, published on 14 January, examined 218.1 million clinical notes from the electronic health records of 1.6 million people receiving care. The findings revealed a notable rise in emoji usage by both healthcare professionals and patients between January 2020 and September 2025.
Across all records analysed, researchers identified 372 unique emojis appearing in 4,162 notes. While this represents a small proportion of total documentation, the upward trend signals a shift in how digital communication is entering clinical environments.
David Hanauer, clinical associate professor of paediatrics and learning health science at Michigan Medicine, explained that the study was initially driven by simple curiosity.
“It was mostly out of interest, just trying to explore if anything was there at all,” Hanauer said. “Our understanding had been that emojis and other symbols are actually not supposed to be used in a medical record, so we were wondering: Were there any there at all, and how often were they being used, and which ones?”
Concerns around clarity and misinterpretation
Despite their increasing use, emojis raise important concerns about clarity in clinical communication. Hanauer highlighted that the ambiguity of many emojis could lead to misunderstanding.
“Most of the concerns that people have is that it’s hard to understand from an emoji what is being conveyed,” Hanauer said. “Maybe a smiley face is pretty obvious to most people, but there’s a lot of different faces with nuances and other symbols. I think there can be a lot of miscommunication, misinterpretation.”
The issue becomes more complex when considering variation in interpretation across different groups of people receiving care and healthcare professionals.
Kim Ford, a health information business systems analyst lead at Michigan Medicine, emphasised how generational differences may influence understanding.
“If you have older patients who may not be familiar with emojis, it’s almost like a foreign language to them,” Ford said. “(For) our younger generation – or those people that have grown up with technology – it’s a second language for them that they understand very well. That’s my biggest concern.”
Accessibility challenges for some people receiving care
Beyond interpretation, accessibility presents another potential barrier. Hanauer noted that small visual symbols may be difficult for some individuals to distinguish clearly, particularly those with visual impairments.
“For older people, having small emojis might actually be hard for them to see and make out, so they might see its face but they can’t tell what the specific expression is,” Hanauer said. “I think we found over 300 different kinds of emojis being used. That’s a lot of different symbols that people would have to understand what they mean.”
This highlights a broader concern that even seemingly simple visual cues may not be universally interpretable or accessible.
Potential implications for patient care
A key concern raised by the study is whether emoji misinterpretation could affect clinical outcomes. While there is currently no direct evidence linking emoji use to adverse outcomes, the possibility remains.
“We hope that doesn’t happen, but I think because of that concern, there’s probably going to be a little bit more oversight,” Hanauer said. “I don’t think we would easily be able to find a circumstance in which there was actually some sort of better or negative outcome from an emoji being misinterpreted.”
The absence of clear evidence does not eliminate the risk, particularly in high-stakes environments where precise communication is essential.
Balancing efficiency with professionalism
Some healthcare professionals recognise potential benefits in using emojis, particularly in reducing communication burden within electronic systems. However, concerns remain about maintaining professionalism and objectivity.
Leah Beel, a medical assistant at American Family Care in Ann Arbor, expressed reservations about their place in formal documentation.
“From my experience, EHRs are used to get quick information and try to communicate with each other in a fast and reliable way,” Beel said. “The only thing I would use is an exclamation point, which, even then, is kind of out there. It’s a good thing that emojis can show enthusiasm or certain reactions, but I also think to a degree – it’s not unprofessional but just someone might take it the wrong way. My perspective on EHR is that you write very objectively.”
In contrast, Elizabeth Rossmann Beel, a paediatric anaesthesiologist at Texas Children’s Hospital, noted that emojis may offer a more efficient way to communicate in certain contexts.
“It’s a way to react to something without putting as much effort into it, or into making that person who’s reading it feel like they need to reply,” Rossmann Beel said. “I think it can cut down a little bit on the burden of replying to and responding to messages in the EHR, which is nice. However, it’s definitely more casual, and so sometimes that’s not the best tone to be setting in a medical record.”
The case for governance and standardisation
Given the growing use of emojis, there is increasing interest in whether formal guidance or regulation should be introduced.
Ford suggested that healthcare organisations may need to consider structured governance around emoji use.
“Maybe emojis are an acceptable means of communication,” Ford said. “The other piece is, should there be a governance process around what emojis can be used? And in what situations? I need to think a little bit about what their structure might look like – what department should be involved in reviewing and approving those, what should be the process to submit an emoji for consideration for use? There’s a lot of pieces to the governance process that need to be figured out there.”
A shift in digital communication within healthcare
The findings from this study reflect a broader evolution in digital communication, where informal elements are beginning to intersect with traditionally formal systems such as electronic health records.
While emojis may offer efficiency and emotional nuance, their integration into clinical documentation raises important questions about clarity, accessibility, professionalism, and patient safety. As their use continues to grow, healthcare systems may need to balance flexibility with standardisation to ensure communication remains precise, inclusive, and clinically appropriate.
Read More
AI Tool Creates ‘Digital Twins’ of Patients to Forecast Future Health
Key Takeaways:
- New DT-GPT model creates virtual patient replicas to predict individual health trajectories with notable accuracy.
- The model outperformed 14 leading machine learning systems and demonstrated effective zero-shot predictions.
- Technology could accelerate drug development and shift healthcare towards more predictive and personalised practice.
Introduction
A new artificial intelligence model capable of generating virtual patient representations and forecasting future health outcomes has been described as a potential breakthrough for clinical research. The system, developed by researchers at the University of Melbourne, uses large language model (LLM) techniques to create personalised digital twins that mirror each individual’s clinical profile.
How the DT-GPT model was developed
The research team trained an existing large language model on three extensive datasets containing thousands of electronic health records. These datasets included information on people living with Alzheimer’s disease, people with non-small cell lung cancer, and people admitted to intensive care units. The aim was to equip the model with sufficient breadth of clinical data to enable it to generate detailed patient-level predictions.
The resulting tool, named DT-GPT, analysed each person’s medical history, such as laboratory values, diagnoses, and treatments. Using this information, it constructed a virtual counterpart for every individual and projected how their condition might evolve under ongoing clinical care.
Predictive performance and validation
Crucially, the model was not shown any actual health outcomes during training. This allowed researchers to rigorously assess the accuracy of its predictions once the model generated forecasts.
Associate Professor Michael Menden, lead researcher, explained the approach:
“For each patient, we created a virtual replica by initialising the model with their individual clinical profile.”
He added:
“For example, we created virtual twins of 35,131 intensive care unit (ICU) patients and accurately predicted what would happen to their magnesium levels, oxygen saturation and their respiratory rate over a 24 hour period, based on their laboratory results from the previous day.”
When benchmarked against 14 state-of-the-art machine learning models, DT-GPT consistently outperformed them in predictive accuracy.
Implications for clinical trials and personalised medicine
Researchers believe the tool has significant implications for the future of clinical trials. Because the model can simulate potential outcomes for large groups of virtual participants, it may help streamline drug development processes by reducing time and cost associated with early-stage testing.
Associate Professor Menden said:
“This technology paves the way for a shift from reactive to predictive and personalised medicine.”
He continued:
“It could enable doctors to anticipate if their patient’s health will deteriorate so they can intervene earlier.
“It could also be used to predict negative side effects of medications, allowing doctors to tailor treatment plans to suit each patient’s unique characteristics and medical history, ultimately increasing the chances of a positive health outcome.”
Conversational interface and handling of complex data
One of DT-GPT’s strengths is its ability to interpret large volumes of complex, unstructured clinical data. The system also includes a conversational interface that functions similarly to a chatbot, enabling clinicians and researchers to query the model directly and explore the reasoning behind specific predictions.
Zero-shot predictions: an advanced capability
Because DT-GPT is based on generative AI, it can also perform zero-shot predictions. These are informed estimates of clinical values that the model has not been explicitly trained to predict.
Associate Professor Menden illustrated this:
“To use an analogy, it’s like asking the model to predict how tall someone will grow without providing the person’s height records and only giving their previous weight and shoe sizes.”
He noted a key finding:
“Our model accurately predicted how lactate dehydrogenase (LDH) levels changed in non-small cell lung cancer patients 13 weeks after they started therapy, despite not training the model for this purpose.
“We compared it to traditional machine learning models, which were specifically trained for 69 clinical variables, including LDH, which we in comparison only educated guessed.
“Very surprisingly, the DT-GPT’s zero-shot predictions, its untrained guesses, were more accurate in 18 percent of cases.”
The study was recently published in NPJ Digital Medicine.
Next steps: expanding to other conditions
The team responsible for developing DT-GPT, in collaboration with the Royal Melbourne Women’s Hospital, have now established the foundation for a new company that will apply digital twin technology to support people living with endometriosis. This work highlights the potential wider applicability of the model across different medical conditions.
Read More
Ambient AI Helps 93% of Doctors Provide Patients with Their “Full Attention”, Sutter Health Study Shows
Key Takeaways:
- Ambient augmented intelligence significantly reduced after-hours documentation and cognitive burden among participating clinicians, with 93 percent reporting they could give patients their full attention.
- Burnout indicators improved, with self-reported after-hours note-taking falling sharply and overall stress scores declining following the pilot’s introduction.
- Despite early challenges around EHR integration and note customisation, clinicians expressed strong enthusiasm for continued use and future development of the technology.
Introduction: Tackling documentation burden with ambient AI
For many clinicians, the administrative workload associated with electronic health record (EHR) systems extends well into the evening, contributing to frustration, diminished work satisfaction and widespread burnout. At Sutter Health in California, leaders have undertaken a substantial effort to determine whether ambient augmented intelligence (AI) could help relieve this pressure and restore time and attention to patient care.
A recent pilot study, published in JAMA Network Open, involved physicians and non-physician providers across the organisation. Participants reported spending less time on after-hours notes, feeling more present with patients during consultations and experiencing early signs of reduced stress. Although limitations remain, the findings suggest that carefully implemented AI-supported documentation could contribute meaningfully to clinician well-being.
National data from the American Medical Association (AMA) illustrate the scale of the problem. Burnout rates among physicians peaked at 62.8 percent in 2021, before falling back to near-2011 levels by 2023. Clinicians remain 82 percent more likely to report burnout than workers in other fields, according to research published in Mayo Clinic Proceedings.
The documentation challenge: A core driver of burnout
The EHR has long been identified as a key contributor to rising workload. Prior research shows that clinicians are “spending two hours of desktop medicine documenting for every hour that they’re spending with patients,” noted Veena Jones, MD, a paediatrician and Sutter Health’s Chief Medical Information Officer, speaking at the 2025 American Conference on Physician Health in Boston.
Sutter Health is a member of the AMA Health System Member Program, which supports health systems with enterprise-level tools designed to strengthen leadership and improve the future of clinical care.
Dr Jones highlighted the cumulative impact of documentation on clinician well-being: “Another national survey showed that about 77 percent of physicians reported that these excessive documentation tasks were leading to longer clinic hours or the need to work from home. Those clinicians who indicated that they had a more favourable view and experience and were highly satisfied with the EHR were less likely to be burned out, which can suggest that changes made to the EHR, particularly through documentation, may be able to provide some relief to this.”
Pilot design: Bringing ambient AI to 100 clinicians
The pilot involved 100 clinicians across multiple specialties and eight medical groups in Northern and Central California. Leaders intentionally recruited a diverse cohort, including primary care clinicians, various specialty clinicians and informatics champions who could model the technology for peers.
Survey findings following the pilot demonstrated significant improvements:
- The proportion of clinicians reporting they spent one hour or less each week on after-hours notes rose from 14 percent to 54 percent.
- The percentage who felt able to give patients their full attention increased from 58 percent to 93 percent.
- Burnout scores dropped from 42 percent to 35 percent.
Cheryl Stults, PhD, senior scientist at the Sutter Health Centre for Health Systems Research, described reductions in cognitive burden: “Regarding task load and cognitive burden, all three of the measures – difficulty accomplishing note writing performance, having to complete notes at a hurried and rush pace, and just the overall mental demand from these tasks – decreased statistically significantly from the pre to the post period.”
The AMA continues to lead efforts to reduce administrative strain through targeted support and system reforms to help clinicians rediscover a greater sense of professional fulfilment.
Early limitations: Integration and customisation gaps
Despite promising outcomes, clinicians identified several challenges during the pilot. These included limited EHR integration, reduced freedom to customise note formats and gaps in specialty-specific templates for physical examinations.
Stults noted: “Despite all of the benefits, there were also some challenges and limitations that they noted from their experience with AI. When our pilot was launched back in April 2024, at the time it was not fully integrated into the EHR. Physicians either had to copy and paste into the EHR or do an additional step to incorporate it into that.”
Since the pilot, full EHR integration has been implemented, resolving one of the most significant issues.
Clinicians also wanted greater flexibility in document structure. As Stults explained: “Additionally, physicians were unhappy that they were unable to customise or format the progress note for future ones, so if they like their note formatted a certain way, they would have to do it every single time – they wanted a way for the AI to remember or to have a level of permanent customisation.” The inclusion of clinicians from a wide range of specialties was intentional, helping ensure templates could be refined more effectively over time.
Participants also sought further functionalities, such as greater accuracy in direct dictation and more precise word-for-word transcription.
Despite these limitations, enthusiasm remained high. As one clinician commented, “I’m very committed to making this work and I really believe that AI will be the way we chart in the future.”
The AMA’s broader work in digital health includes the recent launch of the AMA Centre for Digital Health and AI, designed to ensure clinicians have a strong voice in shaping the use of AI technologies in patient care.
Scaling responsibly: Support over mandates
Following full integration into the EHR, Sutter Health transitioned from the pilot phase to systemwide expansion. Clinicians opted in using a simple self-service form, and most were able to implement the technology after completing two short e-learning modules.
Dr Jones explained: “Part of the uncertainty of knowing how this would go drove us towards a staged monthly implementation where we had our physicians indicate interest with subsequent onboarding. Once we had full EHR integration, we began a self-enrolment process, which was a really simple form. If anyone wants it, they go to our site, they sign up and within a month they will be provisioned.”
Training was streamlined as well. Early analysis suggested that more than two-thirds of clinicians felt confident going live without intensive support. As a result, Sutter Health created a self-guided e-learning module consisting of two seven-minute videos.
Clinical champions remained available to provide at-the-elbow guidance, while a digital academy support team carried out follow-up and troubleshooting.
The AMA’s STEPS Forward webinar, “AI Tools for Documentation: The Newest Member of the Care Team,” provides further insight into how ambient AI can support clinicians and improve care delivery.
Monitoring use and supporting adoption
Sutter Health monitors engagement through monthly utilisation reports. Dr Jones described the organisation’s proactive outreach strategy: “We run monthly reports looking at utilisation and have the team do targeted outreach to those who are not using it to say: Hey, can we help you? And if not, we actually go through a licence repurposing programme.”
This targeted support has helped increase adoption considerably. In March, Sutter Health also became the first organisation to launch a fully integrated inpatient workflow with its ambient AI vendor. This decision came only after the integrated tools demonstrated sufficient maturity to support hospital-based documentation.
As of September, the organisation has been extending the self-service enrolment model across hospitals and emergency departments (EDs).
The shift in clinician demand has been striking. As Dr Jones observed, the usual dynamic of “pushing” new technology has shifted towards clinicians actively requesting access: “The pull versus push has been incredible. In my career, this is one of the most exciting things to be a part of because physicians are pulling for it, and they want it. We have over 1.2 million notes written and that’s increasing at 50,000 a week.”
Read More
Mayo Clinic Launches AI-Powered Nurse Virtual Assistant to Support Frontline Care
Key Takeaways:
- Mayo Clinic nurses have led the design and development of an in-house, AI-powered Nurse Virtual Assistant to streamline access to clinical information.
- The tool consolidates patient summaries, evidence-based guidelines, and clinical policies into one tab within the electronic health record, reducing administrative burden.
- More than 9,600 nurses across Mayo Clinic’s inpatient and emergency departments are now using the system, with ongoing feedback ensuring it continues to evolve with nursing practice.
Streamlining access to critical information
High-quality care depends on timely access to electronic health records, clinical policies, and evidence-based practice guidelines. However, navigating multiple systems to retrieve this information can be time-consuming for nurses, detracting from direct patient care.
To solve this challenge, Mayo Clinic’s Department of Nursing led a multidisciplinary initiative to create Nurse Virtual Assistant – a generative artificial intelligence (AI) tool developed entirely in-house by nurses for nurses. Integrated into Mayo Clinic’s electronic health record (EHR) system, the tool presents essential information within a single tab, making it easier to retrieve and act upon.
Nurses can view a curated, nurse-specific patient summary and access direct links to key evidence-based resources, including Lippincott procedures, IV administration guidelines, and Mayo Clinic’s clinical policy library – all in one place.
This streamlined interface allows nurses to spend less time searching and more time focusing on what matters most: the person receiving care.
Augmenting – not replacing – human connection
Mayo Clinic emphasises that Nurse Virtual Assistant is designed to support, rather than replace, the expertise and human presence that nurses bring to clinical practice.
“It is an amazing tool,” says Nick Flynn, a registered nurse in the Emergency Department at Mayo Clinic Hospital in Arizona. Flynn highlights the value of consolidated patient data from inpatient stays, outpatient visits, and phone calls:
“You have easy access to a history of their illness, and that is available just moments after they arrive.”
Nurse-driven innovation from concept to rollout
The project began in 2024 as part of Mayo Clinic’s strategy to ease administrative pressures in a rapidly digitising healthcare environment. Crucially, nurses were involved at every stage – from conceptualisation to design and testing – ensuring the tool meets real-world clinical needs.
Early-access users played an active role in shaping the system’s features, providing feedback that directly informed improvements.
Brendon Bloomfield, a registered nurse in Psychiatric Acute Care at Mayo Clinic Hospital – Rochester, explains:
“To see a concept I was passionate about, AI-enhanced communication, actually get built – and to be invited to help shape it – reinforces that frontline nurses’ voices matter and that we have the power to influence the future of care.”
The solution underwent a research study approved by an Institutional Review Board before scaling to more than 9,600 nurses across inpatient and emergency department units.
Evolving with nursing practice
Nurse Virtual Assistant has been released as a Minimum Lovable Product – a version that not only solves an immediate problem but is designed to be engaging and impactful for end users. Nurses can submit feedback directly through the tool, allowing the solution to continuously evolve with frontline input.
Enhancements already implemented include improved search result accuracy, refined content layouts, and new functionality based on user suggestions.
Privacy, security, and compliance at the core
Built to the highest standards of data privacy, the Nurse Virtual Assistant is a patent-pending solution developed in full compliance with HIPAA regulations and other applicable requirements. This ensures that patient information remains secure while enabling timely and efficient access for clinical teams.
Shaping the future of nursing care
Mayo Clinic’s Chief Nursing Officer, Ryannon Frederick, sees the innovation as a milestone in supporting nursing practice:
“Nurse Virtual Assistant is an example of how Mayo Clinic nurses are driving innovation and shaping the future of care. By reducing administrative burden, we allow nurses to focus on the most important part of their work: caring for patients with skill, compassion and presence.”
Read More
Vanderbilt Researchers Use AI to Address Gaps in Long-Term Obesity Care
Key Takeaways:
- A $1 million Eli Lilly grant will fund a two-year Vanderbilt University Medical Center (VUMC) project using artificial intelligence (AI) to address gaps in obesity care.
- The initiative will analyse electronic health records (EHRs), survey patients and clinicians, and build a multi-agent AI system to develop evidence-based strategies for improving long-term engagement.
- A patient-facing mobile application will be designed and piloted in VUMC obesity clinics to support shared decision-making and sustained weight management.
Major investment in addressing gaps in care
Vanderbilt University Medical Center (VUMC) has secured a $1 million grant from Eli Lilly and Company to fund a two-year research project aimed at improving continuity of care for people living with obesity. The initiative seeks to understand why many individuals discontinue treatment and to create scalable solutions to help them stay engaged in long-term care.
“Obesity is a chronic, relapsing condition that requires ongoing management, yet too often it is treated episodically because of barriers like delayed medication access,” explained You Chen, PhD, Associate Professor of Biomedical Informatics and the project’s Principal Investigator for informatics and technology. “We’re combining data-driven insights, stakeholder input and multi-agent AI to understand where continuity breaks down and to design evidence-based interventions that keep patients engaged.”
Data-driven insights and stakeholder engagement
In its first year, the research team will analyse VUMC’s electronic health records to identify patterns distinguishing people who remain in continuous follow-up from those who disengage. Patient and clinician surveys will be conducted to capture real-world barriers to care, including logistical, financial and psychological challenges.
The findings will be integrated into a multi-agent AI system, featuring simulated physician, nurse and dietitian agents. This system will generate and prioritise strategies for maintaining engagement, which will then be reviewed by panels of clinicians, informaticians and patient representatives.
Patient-facing app to support engagement
The second year of the project will focus on designing and piloting a mobile application to be used in VUMC obesity clinics. This app is intended to help patients view and interpret their own health data, complete pre-visit tasks, and communicate more effectively with their care teams.
“By helping patients view and interpret their own data, complete previsit tasks, and communicate more effectively with care teams, the app will aim to strengthen shared decision-making and sustain engagement over time,” said Chen.
Clinical leadership and broader impact
The project’s clinical lead is Gitanjali Srivastava, MD, Professor of Medicine in the Division of Diabetes, Endocrinology and Metabolism.
“Medicine has evolved, and we need to adapt to new technological advances while catering to patient needs,” Srivastava stated. “It’s about designing practical tools and processes that fit naturally into patients’ lives and clinicians’ workflows, ultimately supporting healthier weight management over time.”
Chen emphasised that the project is intended to be scalable across health systems. The researchers believe that the human–AI collaborative approach developed through this project could serve as a reproducible framework for improving continuity of care for other chronic conditions that require long-term management.
Read More