Illustration by Adobe Stock/Diana Duren
Vanderbilt Health researchers are at the forefront of efforts to integrate artificial intelligence into health care and biomedical research.
They’ve published recent studies demonstrating how AI can reduce clinical alert fatigue, accelerate drug repurposing, improve patient communication, diagnose rare diseases, predict treatment responses, and generate therapeutic antibodies.
“Part of our mission at Vanderbilt is to develop and implement innovative AI methods, frameworks and tools to advance data-driven, precision health care and biomedical discovery, all while remaining committed to ethical and legal guidelines,” said Peter Embí, MD, MS, professor of Biomedical Informatics and co-director of the department’s ADVANCE Center (AI Discovery and Vigilance to Accelerate Innovation and Clinical Excellence). “Vanderbilt’s collaborative teams of clinicians, data scientists and basic science researchers are uniquely positioned to apply generative AI, predictive modeling, machine learning and advanced data analytics to our clinical and operational systems, as well as to advance the foundations of biomedicine.
“As an institution, we are well connected to and often play a leadership role in AI-related networks and initiatives that help inform our strategy and path forward in this rapidly developing field. As part of this transformation, Vanderbilt teams of scientists and operational experts gather critical data to better track and ensure AI’s positive impacts on health care, informing both our practice and the advancement of this work nationally and internationally,” said Embí, who holds the Directorship in Biomedical Informatics.
The following is a collection of AI studies and funding awards from the past two years, spanning clinical decision support, computational biology, mental health screening, cancer care, and infectious disease prevention. This work reflects millions of dollars in federal funding and involves collaborations across multiple disciplines aimed at a common goal of using AI to improve patient outcomes and advance medical knowledge.
AI can help patients craft better portal messages
Reporting in the Journal of the American Medical Informatics Association, Siru Liu, PhD, Adam Wright, PhD, and colleagues explored using large language models to help patients craft more effective messages to health care providers through patient portals.
In a blinded test, their custom-trained model, CLAIR, produced follow-up questions for patients with similar clarity and concision and higher utility than actual follow-up questions written by care team members. The idea is that before patients hit “send,” AI could prompt them to clarify their portal messages.
Such a tool could streamline communication and improve care efficiency by reducing back-and-forth messaging. The team previously established that AI is better than doctors at responding to patient messages. They plan to evaluate the clinical impact of implementing AI-guided patient messaging.
AI flags suicide risk in neurology clinics to spur conversations
Reporting in JAMA Network Open, Colin Walsh, MD, MA, and colleagues tested whether their machine learning-derived Vanderbilt Suicide Attempt and Ideation Likelihood model (VSAIL) could effectively prompt doctors to screen patients for suicide risk. At three neurology clinics, the system flagged 8% of arriving patients as having relatively high risk for a suicide attempt in the next 30 days. Interruptive alerts triggered by the model led doctors to conduct suicide risk assessments 42% of the time, compared to just 4% with passive alerts.
“Most people who die by suicide have seen a health care provider in the year before their death, often for reasons unrelated to mental health,” Walsh said. “But universal screening isn’t practical in every setting. We developed VSAIL to help identify high-risk patients and prompt focused screening conversations.”
AI recruited to lower blood clots
A yearlong randomized clinical trial involving all Vanderbilt Health adult inpatients is testing whether AI can help reduce blood clots that form inside blood vessels during hospitalization, or hospital-acquired venous thromboembolism (HA-VTE). A machine learning-derived risk model analyzes 25 variables commonly found in electronic health records, including vital signs, laboratory results, medical history and procedures like central line placement. In validation studies, the model correctly ranked patients by risk 89% of the time — considered excellent performance as prognostic models go.
“Despite decades of research and numerous risk prediction tools, preventing HA-VTE remains a challenge as even a single case might do great harm,” said the trial’s lead investigator, Colin Walsh, MD, MA. “Our AI-driven approach doesn’t require doctors to manually calculate risk scores or enter any data. Instead, it works quietly in the background, analyzing patient data in real time and alerting clinicians only when action is needed.” Results are expected in 2027.
When AI writes better than doctors
Reporting in the Journal of the American Medical Informatics Association, Siru Liu, PhD, Adam Wright, PhD, and colleagues found ChatGPT-3.5 and ChatGPT-4 outperformed doctors in answering patient questions sent via patient portals.
In a blinded test, four primary care doctors rated actual responses from doctors and responses from AI programs to questions about bladder infections, sleep issues, back pain prescriptions, flu symptoms, blood in stools and COVID-19. Judging involved four categories: empathy, accuracy, usefulness and responsiveness.
The intent behind the study is to develop AI to write first-draft responses that doctors would use to speed their work. Primary care doctors typically spend 1.5 hours per day processing patient messages.
Large language models show phenotyping promise
Wei-Qi Wei, MD, PhD, Chao Yan, PhD, and colleagues reported in the Journal of the American Medical Informatics Association that large language models can help generate electronic health record phenotyping algorithms. They tested ChatGPT-4, ChatGPT-3.5, Claude 2 and Bard (now Gemini) for generating algorithms for Type 2 diabetes, dementia and hypothyroidism. Three experts evaluated the algorithms, finding ChatGPT-4 and -3.5 significantly outperformed Claude 2 and Bard.
“Developing EHR phenotypes demands substantial informatics and clinical knowledge. It’s an intricate process that limits the pace of research,” Wei said. Testing on data from over 80,000 patients showed mixed performance compared to gold-standard expert-developed algorithms.
“These AI models show exciting capabilities, but they’re not yet ready to generate expert-level phenotyping algorithms, certainly not right out of the box,” Wei said. “We believe they can help jump-start the process, allowing experts to focus more on fine-tuning rather than starting from scratch.”
Expanding AI-powered chart abstraction of vital information
Daniel Fabbri, PhD, and Christine Micheel, PhD, secured a one-year, $2 million extension of their Advanced Research Projects Agency for Health award for an AI-guided platform for extracting and organizing critical information from unstructured clinical notes and reports in medical records.
The platform has already been spun into a commercial venture called Brim Analytics and is being used around the country to support clinical research, cancer registries and health initiatives. Over the past year, more than 120 research and clinical teams at Vanderbilt Health used Brim to perform chart abstraction for applications from surgical planning and cancer registries to orthopaedic research.
“Brim has quickly become part of our intelligence system infrastructure at Vanderbilt,” said Health IT leader Neal Patel, MD, MPH.
AI platform maps tissues in 3D
In Nature Methods, Tae Hyun Hwang, PhD, and colleagues published validation of iSCALE, an AI platform for imaging gene expression. The platform produces high-resolution maps of cellular landscapes across tissue sections.
The demonstration used postmortem tissue samples from patients with gastric cancer and multiple sclerosis. The authors wrote that their findings show “the utility of iSCALE in analyzing large tissues by enabling unbiased annotation, resolving cell type composition, mapping cellular microenvironments and revealing spatial features beyond the reach of standard spatial transcriptomic analysis.”
“This work fundamentally reimagines how we can scale spatial biology to clinically relevant samples,” Hwang said. “Instead of testing thousands of points across a slide, we can now learn from just a few regions and generate a molecular map of the whole tissue. That changes the equation for translational science and diagnostics.”
Predicting who benefits from cancer immunotherapy
In a research partnership with GE HealthCare, Travis Osterman, DO, MS, and colleagues published results in the Journal of Clinical Oncology Clinical Cancer Informatics showing that AI models predicted patient responses to immunotherapy with 70% to 80% accuracy. The models use routinely collected electronic health record data like diagnosis codes and medications.
“We focused primarily on this routinely collected structured data to build predictive models with the goal that these models would be able to be implemented in any clinical setting,” Osterman said.
GE HealthCare is evaluating plans to commercialize the models for pharmaceutical drug development and clinical support. The researchers said the methodology has potential for use in other areas such as neurology or cardiology.
Making health chatbots safe and empathetic
Susannah Rose, MSSW, PhD, and Zhijun Yin, PhD, received up to $7.3 million from the Advanced Research Projects Agency for Health to build the Vanderbilt Chatbot Accuracy and Reliability Evaluation System (V-CARES). Using mental health as a demonstration case, the system will focus on detecting hallucinations, omissions and misaligned values in AI-generated responses on critical health topics.
“We chose screening and treatment for major depression and generalized anxiety disorder because, from a safety and reliability standpoint, these chatbots pose a number of lingering challenges and unresolved questions,” Rose said.
The project will combine human expertise with advanced computational techniques.
“We will pursue a novel multiexpert ensemble learning framework,” Yin said, “integrating various AI models and human expertise to achieve accurate detection of potential issues in chatbot responses.”
ChatGPT finds drug candidates for Alzheimer’s
Reporting in npj Digital Medicine, Chao Yan, PhD, Monika Grabowska, Wei-Qi Wei, MD, PhD, and colleagues used ChatGPT to identify existing drugs that might help patients with Alzheimer’s disease. They prompted the chatbot to suggest and then confirm 20 drugs for Alzheimer’s disease, repeating this exercise 10 times.
Analyzing medical records from Vanderbilt Health and the All of Us Research Program, they found three of the suggested drugs were associated with lower Alzheimer’s risk across both datasets: losartan with 24% reduced risk, metformin with 33% reduced risk, and simvastatin with 16% reduced risk.
“LLMs like ChatGPT speedily accomplish a form of extensive literature review, which has become infeasible for humans to perform alone,” Wei said.
Designing novel antibodies against emerging viral threats
Ivelin Georgiev, PhD, Perry Wasdin, PhD, and colleagues reported in Cell that a so-called protein language model could design functional human antibodies against existing and emerging viral threats. Training their generative monoclonal antibody model on previously characterized antibodies against a known avian influenza strain, the researchers generated antibodies against a related strain.
“This study is an important early milestone toward our ultimate goal — using computers to efficiently and effectively design novel biologics from scratch and translate them into the clinic,” Georgiev said. “Such approaches will have significant positive impact on public health and can be applied to a broad range of diseases, including cancer, autoimmunity, neurological diseases and many others.”
Georgiev received an award of up to $30 million from the Advanced Research Projects Agency for Health to develop AI technology for creating novel antibodies.
Dual technologies for cancer care
Tae Hyun Hwang, PhD, helped develop two AI technologies for improving cancer care. A framework described in Nature Communications integrates deep learning with a microscopic 3D imaging technique called holotomography to generate digitally “stained” images from tissue samples.
The noninvasive technique preserves tissue integrity, ensuring compatibility with downstream molecular assays, and has potential beyond cancer diagnostics.
“This technology fundamentally redefines how we visualize and analyze tissue architecture, moving from traditional two-dimensional views to full 3D microenvironment mapping at the subcellular level,” Hwang said.
A second technology, MSI-SEER, identifies patients who will benefit from immunotherapy that might otherwise be missed. As described in npj Digital Medicine, the tool uses standard pathology slides to better predict microsatellite instability, a condition where cancer cells have numerous errors in short, repeated DNA sequences.
Mining mortality data from the web
Reporting in the Journal of Medical Internet Research, Mohammed Ali Al-Garadi, PhD, Ruth Reeves, PhD, and colleagues used natural language processing to collect mortality information from crowdfunding platforms, web-based obituaries and memorial websites.
Quick, low-cost collection and processing of this information and its linkage with patient records could aid large-scale health research, medical device safety monitoring, and timeliness of public health measures.
Some 8.1 million retrieved documents were analyzed. An open-source large language model (Meta’s Llama-13) performed on par with nurses trained as research assistants in understanding and annotating raw information.
Using AI to keep patients receiving obesity treatment engaged
You Chen, PhD, and Gitanjali Srivastava, MD, received $1 million from Lilly to study and address gaps in obesity care. The project will analyze electronic health records, survey patients and clinicians, and use AI to identify why many patients discontinue obesity treatment.
“Obesity is a chronic, relapsing condition that requires ongoing management, yet too often it’s treated episodically,” Chen said.
Findings will feed a multiagent AI system — physician, nurse, dietitian — that will generate ideas for consideration by panels of clinicians, informaticians and patient representatives. In the second year, the team will design and implement a patient-facing mobile app.
“Medicine has evolved, and we need to adapt to new technological advances while catering to patient needs,” Srivastava said. The researchers said their human-AI collaborative methodology could extend beyond obesity care.
Creating synthetic hiv patients for research
Bryan Shepherd, PhD, and Bradley Malin, PhD, received a five-year, $4 million grant from the National Institutes of Health to create hundreds of thousands of simulated HIV patients to aid longitudinal observational studies. Synthetic data created by the team will be made public.
“We think simulated data can greatly benefit HIV research, particularly in international settings where data sharing is becoming more complicated. … The rate of discovery is being impeded by sensitivities around HIV and legitimate privacy concerns,” Shepherd said.
Using generative AI, the team will attempt to mimic patients from two large multinational HIV research cohorts numbering more than 500,000 people living with HIV.
“Sharing observational datasets is essential to enabling new hypotheses leading to potential cures,” Malin said. “Findings from our lab and others point to patient simulation as an alternative to typical patient de-identification methods that reduce data fidelity and research utility.”
Shortening the diagnostic odyssey for rare diseases
Cathy Shyr, PhD, Rizwan Hamid, MD, PhD, and colleagues reported in JAMA Network Open that large language models (LLMs) successfully identified diagnoses for patients in the Undiagnosed Diseases Network.
“In some cases, the diagnostic odyssey for patients with rare disorders — the time from symptom onset until diagnosis — lasts for more than 10 years,” Hamid said.
They assessed whether LLMs can identify final diagnoses for UDN patients based on clinical summaries. Compared to the historical human clinical review rate of 5.6%, the LLMs achieved diagnostic rates of 13.3% (ChatGPT 4o) and 10.0% (Llama 3.1 8B). Cost per case ranged from zero to 3 cents depending on the model.