Acute gastrointestinal (GI) bleeding is the most common GI diagnosis requiring hospitalization in the United States. Patients with acute GI bleeding either vomit up blood or have blood in their stool and must be evaluated in the hospital.
“When a patient arrives at the hospital with acute GI bleeding, physicians need to quickly figure out answers to three questions: Why is the patient bleeding? What supportive care can help them get through this episode of bleeding? How can I stop the bleeding long term?” said Dennis Shung, MD, PhD, assistant professor of medicine (digestive diseases, biomedical informatics and data science) and director of digital health (digestive diseases).
While all acute GI bleeding is abnormal, some patients are very low-risk and may not need to remain in the hospital. Current guidelines recommend physicians use clinical risk scores, like the Glasgow Blatchford Score (GBS) or the Oakland Score, to help determine which patients may need medical intervention, such as a blood transfusion or endoscopic procedure, and which patients may be safely discharged from the hospital.
“We don’t always know the location of the bleeding source,” said Loren Laine, MD, professor of medicine (digestive diseases) and chief of Yale Digestive Diseases. “Acute GI bleeding can stem from various causes, including ulcers, tumors, or inflammation anywhere throughout the GI tract. This limits the usefulness of the current clinical risk scores, which require the physician to know the location of the bleeding”.
Yale researchers are developing a machine-learning risk model that uses electronic health record (EHR) data to provide an initial patient assessment based on symptoms rather than where the bleeding is thought to be.
“Caring for patients in a hospital is complex and fast-paced. Physicians shouldn’t need to spend time punching in numbers that already exist within the EHR,” said Shung. “Our model extracts clinically relevant, structured data fields – like labs, medical history, medications, age, vital signs, and Glasgow Coma Scale scores – from the EHR to give a real-time risk assessment of each individual patient.”
In a new paper published in Gastroenterology, Shung, Laine, and colleagues evaluated the machine learning risk model against existing clinical risk scores to determine the most accurate approach. The study showed that the EHR-based machine learning model performs better than existing clinical risk scores and identifies more very low-risk patients who can safely leave the hospital.
“Many patients who are very low-risk do not want – or need – to remain in the hospital,” said Laine, co-senior author of the paper. “If we can quickly and more accurately identify very low-risk patients, we can reduce patient inconvenience and reduce health care costs – all without harm to the patient.”
The model outlined in the paper uses data fields from Epic, the most widely used EHR globally, which offers the potential for health systems across the world to tailor the data fields for their particular patient populations and unique needs.
The paper also found that while the machine learning algorithm performed better, both the GBS and Oakland scores performed well. Shung said this finding has led him and his team to consider designing a clinical decision support tool that shows the GBS score, Oakland score, and a machine learning score so that they can see three views on the patient.
“Machine learning systems are complex, and, at least for now, they may not have the trust that the other validated clinical risk scores do,” said Shung, who is the first author of the paper. “We don’t need to throw out the clinical risk scores, but we can build upon them and add new information to create better clinical decision support tools. This could be the future of risk stratification.”
Shung, alongside other Yale researchers, continues to investigate how doctors and medical students interact with artificial intelligence (AI) tools, including GutGPT, a generative AI that allows them to ask questions to treat hypothetical patients with GI bleeding.
“The lesson I take away from my research is that it’s not just about a technology and how it performs,” said Shung. “We’re human, and we all bring our own experiences to the human-computer interaction. We need to figure out how to integrate these advanced technologies into a human, complex care delivery system. It’s a socio-technical challenge. We're excited about exploring in all these different ways.”
“Validation of an Electronic Health Record-Based Machine Learning Model Compared to Clinical Risk Scores for Gastrointestinal Bleeding” was published in Gastroenterology. Other authors of the paper include Colleen Chan, PhD; Kisung You, PhD; Neil S. Zheng, MD; Shinpei Nakamura; Theo Saarinen; Michael Simonov, MD, lecturer in biostatistics (health informatics); Darrick K. Li, MD, PhD, assistant clinical professor (digestive diseases); Cynthia Tsay, MD; Yuki Kawamura, MD, PhD, associate research scientist; Matthew Shen; Allen Hsiao, MD, professor of pediatrics (emergency medicine) and professor of emergency medicine; and Jasjeet Sekhon, Eugene Meyer Professor of Political Science and Professor of Statistics and Data Science (co-senior author).
Since forming one of the nation’s first sections of hepatology more than 75 years ago and then gastroenterology nearly 70 years ago, Yale School of Medicine’s Section of Digestive Diseases has had an enduring impact on research and clinical care in gastrointestinal and liver disorders. To learn more about their work, visit Internal Medicine: Digestive Diseases.