image: Study procedure flow chart. After patients with bloodstream infection (BSI) were identified from microbiology databases, 27 patients' characteristics were extracted from electronic health records. After the dimension of those variables was reduced with Uniform Manifold Approxi mation and Projection (UMAP), k-means++ was used to identify clusters. We identified 3 clinically distinct clusters that were identified. Among those clusters, patients were divided into solid organ transplant (SOT) and non-SOT groups. Clinical characteristics and outcomes were evaluated.
Credit: American Journal of Transplantation
Bloodstream infections (BSI) can turn deadly fast, particularly for patients with weakened immune systems. A new study from Houston Methodist Research Institute finds that artificial intelligence can assist clinicians in identifying previously unseen patterns of infection in patients.
Led by Masayuki Nigo, M.D., associate professor in the Department of Medicine at Houston Methodist, the study was published in the American Journal of Transplantation and used an unsupervised machine learning model to identify three major clusters of BSI patients. Analyzing data from more than 15,000 patients, the researchers formed cluster characteristics based on clinical data obtained within the first 48 hours of BSI diagnosis.
“This study is important because it demonstrates that patients with bloodstream infections, including solid organ transplant recipients, are not clinically uniform despite sharing the same diagnosis,” Nigo said. “Using routinely collected data from the first 48 hours of infection, we identified three distinct clinical patterns based on patient characteristics, illness severity and need for organ support through a machine learning–based clustering approach.”
The highest‑risk group captured older, predominantly male patients who required more ventilator and vasopressor support, as well as transplant patients. These individuals are especially susceptible to infections due to their weakened immune systems, with one in 10 experiencing an infection within the first year following their transplant. Mortality rates can be as high as 60%.
In contrast, the two other groups of patients shared similar clinical profiles with only modest differences in their initial characteristics. However, outcomes varied greatly – one group showed patients had milder symptoms, while the other group showed higher severity and higher patient mortality.
“Our model turns routine early data into a risk map clinicians can use immediately,” said Stefano Casarin, Ph.D., assistant professor in the Center for Precision Surgery at Houston Methodist Research Institute. “This gives us a new way to understand and predict how sick a patient might become. If we can identify high‑risk patients sooner, we can act sooner.”
Nigo said that the next steps will involve validating the findings in external health care systems to confirm reproducibility and further studies to determine how the methodology can be improved to foster better clinical decision-making and patient outcomes.
Other collaborators in the study are Max Adelman, James Kurian, Jiaqiong Xu, David Hsu, Aarjav Sanghvi, Stephen Jones, Ashton Connor, Ahmed Gaber, Mark Ghobrial and Cesar Arias from Houston Methodist.
The study was supported by funding from the National Institutes of Health’s National Institute of Allergy and Infectious Diseases and the Houston Methodist Academic Institute.
Journal
American Journal of Transplantation