AI Algorithm Predicts Transfusion Need in Acute GI Bleeds

Medscape

2 days ago

SAN DIEGO — A novel generative artificial intelligence (AI) framework known as trajectory flow matching (TFM) can predict the need for red blood cell transfusion and mortality risk in intensive care unit (ICU) patients with acute gastrointestinal (GI) bleeding, researchers reported at Digestive Disease Week (DDW) 2025.
Acute GI bleeding is the most common cause of digestive disease–related hospitalization, with an estimated 500,000 hospital admissions annually. It's known that predicting the need for red blood cell transfusion in the first 24 hours may improve resuscitation and decrease both morbidity and mortality.
However, an existing clinical score known as the Rockall Score does not perform well for predicting mortality, Xi (Nicole) Zhang, an MD-PhD student at McGill University, Montreal, Quebec, Canada, told attendees at DDW. With an area under the curve of 0.65-0.75, better prediction is needed, Zhang said, whose coresearchers included Dennis Shung, MD, MHS, PhD, assistant professor of medicine and director of Applied Artificial Intelligence at Yale University School of Medicine, New Haven, Connecticut.
'We'd like to predict multiple outcomes in addition to mortality,' said Zhang, who is also a student at the Mila-Quebec Artificial Intelligence Institute.
As a result, the researchers turned to the TFM approach, applying it to ICU patients with acute GI bleeding to predict both the need for transfusion and in-hospital mortality risk. The all-cause mortality rate is up to 11%, according to a 2020 study by James Y. W. Lau, MD, and colleagues. The rebleeding rate of nonvariceal upper GI bleeds is up to 10.4%. Zhang said the rebleeding rate for variceal upper gastrointestinal bleeding is up to 65%.
The AI method the researchers used outperformed a standard deep learning model at predicting the need for transfusion and estimating mortality risk.
Defining the AI Framework
'Probabilistic flow matching is a class of generative artificial intelligence that learns how a simple distribution becomes a more complex distribution with ordinary differential equations,' Zhang told Medscape Medical News. 'For example, if you had a few lines and shapes you could learn how it could become a detailed portrait of a face. In our case, we start with a few blood pressure and heart rate measurements and learn the pattern of blood pressures and heart rates over time, particularly if they reflect clinical deterioration with hemodynamic instability.'
Another way to think about the underlying algorithm, Zhang said, is to think about a river with boats where the river flow determines where the boats end up. 'We are trying to direct the boat to the correct dock by adjusting the flow of water in the canal. In this case we are mapping the distribution with the first few data points to the distribution with the entire patient trajectory.'
The information gained, she said, could be helpful in timing endoscopic evaluation or allocating red blood cell products for emergent transfusion.
Study Details
The researchers evaluated a cohort of 2602 patients admitted to the ICU, identified from the publicly available MIMIC-III database. They divided the patients into a training set of 2342 patients and an internal validation set of 260 patients. Input variables were severe liver disease comorbidity, administration of vasopressor medications, mean arterial blood pressure, and heart rate over the first 24 hours.
Excluded was hemoglobin, since the point was to test the trajectory of hemodynamic parameters independent of hemoglobin thresholds used to guide red blood cell transfusion.
The outcome measures were administration of packed red blood cell transfusion within 24 hours and all-cause hospital mortality.
The TFM was more accurate than a standard deep learning model in predicting red blood cell transfusion, with an accuracy of 93.6% vs 43.2%; P ≤ .001. It was also more accurate at predicting all-cause in-hospital mortality, with an accuracy of 89.5% vs 42.5%, P = .01.
The researchers concluded that the TFM approach was able to predict the hemodynamic trajectories of patients with acute GI bleeding defined as deviation and outperformed the baseline from the measured mean arterial pressure and heart rate.
Expert Perspective
'This is an exciting proof-of-concept study that shows generative AI methods may be applied to complex datasets in order to improve on our current predictive models and improve patient care,' said Jeremy Glissen Brown, MD, MSc, an assistant professor of medicine and a practicing gastroenterologist at Duke University who has published research on the use of AI in clinical practice. He reviewed the study for Medscape Medical News but was not involved in the research.
'Future work will likely look into the implementation of a version of this model on real-time data.'
He added: 'We are at an exciting inflection point in predictive models within GI and clinical medicine. Predictive models based on deep learning and generative AI hold the promise of improving how we predict and treat disease states, but the excitement being generated with studies such as this needs to be balanced with the trade-offs inherent to the current paradigm of deep learning and generative models compared to more traditional regression-based models. These include many of the same 'black box' explainability questions that have risen in the age of convolutional neural networks as well as some method-specific questions due to the continuous and implicit nature of TFM.'
Elaborating on that, Glissen Brown said: 'TFM, like many deep learning techniques, raises concerns about explainability that we've long seen with convolutional neural networks — the 'black box' problem, where it's difficult to interpret exactly how and why the model arrives at a particular decision. But TFM also introduces unique challenges due to its continuous and implicit formulation. Since it often learns flows without explicitly defining intermediate representations or steps, it can be harder to trace the logic or pathways it uses to connect inputs to outputs. This makes standard interpretability tools less effective and calls for new techniques tailored to these continuous architectures.'
'This approach could have a real clinical impact,' said Robert Hirten, MD, associate professor of medicine and artificial intelligence, Icahn School of Medicine at Mount Sinai, New York City, who also reviewed the study. 'Accurately predicting transfusion needs and mortality risk in real time could support earlier, more targeted interventions for high-risk patients. While these findings still need to be validated in prospective studies, it could enhance ICU decision-making and resource allocation.'
'For the practicing gastroenterologist, we envision this system could help them figure out when to perform endoscopy in a patient admitted with acute gastrointestinal bleeding in the ICU at very high risk of exsanguination,' Zhang told Medscape Medical News.
The approach, the researchers said, will be useful in identifying unique patient characteristics, make possible the identification of high-risk patients and lead to more personalized medicine.
Hirten, Zhang, and Shung had no disclosures. Glissen Brown reported consulting relationships with Medtronic, OdinVision, Doximity, and Olympus.