Dr. Alison Fohner, an assistant professor in the department of epidemiology, and a team of researchers at Kaiser Permanente in California recently published a data mining tool that uses a machine-learning algorithm to classify patients with sepsis. The outcomes of this study could help develop more targeted therapies against sepsis.
Sepsis is an overwhelming immune response to an infection that is triggered by pathogens. Immune cells are released into the blood to fight the infection, which can lead to blood clots and leaky blood vessels. It is a major public health issue that impacts 1.7 million adults in the United States and kills 270,000. Despite heterogeneity in sepsis, physicians typically use a “one-size-fits-all” treatment of antibiotics and fluids that do not account for patient differences.
“The ways we have now to [classify] these people are very basic and do not work well,” Fohner said. “For example, people with organ failure are grouped together, but each within that group is different.”
Fohner received her Ph.D. from UW’s genetics program and completed a clinical data science fellowship at Kaiser. She is interested in research that applies the differences in patient health outcomes to improve targeted treatments.
“[In] so much of medical research we try to get similar patients,” Fohner said. “If we can use real people and medical records, we can work with tons more people because we do not have to identify 100 people who all look the same.”
For this project, Fohner adapted a text-mining tool that used a machine-learning algorithm to search for trends in medications and procedures administered within the first 24 hours of hospitalization to 29,543 septic patients. This data was extracted from the electronic health record (EHR), a digital patient record that provides physicians with patient summaries but was designed for insurance billing. Each medication, diagnosis, and procedure in the EHR has a code that allows insurance companies to effectively charge for administered services.
Fohner extracted codes for medications, orders, and procedures administered to the study patients.
“[We cared] more about the post-sepsis care,” Fohner said. “We are trying to figure out how these patients are treated to identify who is doing well and why.”
The data was organized to account for the frequency of certain billing codes per patient record. Subsequently, a machine learning algorithm took all patient data and generated 42 topics that were combinations of EHR codes that commonly occur together. The algorithm also identified dominant topics per patient. Creating such subsets of patient types will help researchers create more unique treatment plans based on pre-existing disease patterns.
Despite creating only 42 labels from such comprehensive data, Fohner believes that it is a non-ideal metric to categorize patients.
“It shows how complex and how messy treatment for a condition like this can be, if we are trying to make progress on it,” Fohner said.
After assigning each patient representative labels, Fohner could analyze the associations between different healthcare factors such as pre-existing conditions and medications that contributed to an outcome measure of mortality.
“If you had someone who died, you can weight these features to determine which is more likely to lead to death,” Fohner explained.
The study also evaluated the time at which the first antibiotic was administered and the amount of fluids given to patients. With this study, Fohner wants to improve the identification of patients that need serious care right away. Those who display obvious signs of sepsis are administered antibiotics immediately, but those who do not can have different health outcomes.
By studying a collection of variables, Fohner hopes that clinicians can move past binary tree comparisons of healthcare measures. Fohner illustrates this through an example of two patients, one with heart disease and the other without. Patients with heart disease are administered less fluid because their condition causes fluid build-up.
“If we just compare that to someone who got a lot of fluid right away, then the patient with heart disease received worse care,” she said. “However, [the patient with heart disease] just has more medical complexity and did not get worse care.”
The outputs of this research project feed into precision care, especially for researchers who are trying to develop better therapies. They will not only save time and resources, but they can also determine what mechanisms play a role in patient healthcare. Although there is a current push in machine learning and artificial intelligence as prediction tools in healthcare, Fohner is still cautious.
“I think the barrier to changing care is huge,” she said. “You have to be very certain that what you are proposing will have a measurable impact. You are going to help more people without hurting many.”
Fohner published her code alongside the paper for others to use. By using such automated processes, clinicians can get past the barrier of reading medical notes and create more complex and comprehensive patient profiles based on outcomes. With the vast amounts of data available in the EHR, researchers can guide precision treatments for not only sepsis but other more heterogeneous diseases.
Reach writer Vidhi Singh at email@example.com. Twitter: @vidhisvida
Like what you’re reading? Support high-quality student journalism by donating here.