Optimizing the timing of sepsis treatment with a machine learning model

Machine Learning

A new machine learning model that estimates optimal treatment timing for sepsis could pave the way for supporting tools that help physicians individualize treatment decisions at the patient’s bedside, researchers say. increase.

In a paper published today (April 6, 2023) nature machine intelligencescientists at Ohio State University describe a new model that uses artificial intelligence to address the complex problem of when to give antibiotics to patients with suspected sepsis.

Time is of the essence, as sepsis, the body’s overwhelming response to infection, can quickly lead to organ failure. Still, its symptoms – fever, low blood pressure, increased heart rate, breathing problems – can look like many other conditions. Federal guidelines recommend broad-spectrum antibiotics as the first line of defense. Seeking immediate medical attention. This is usually a strategy that needs to be acted upon before obtaining cultures confirming bacterial infection from the lab.

Ping Chan

This model was designed with these uncertainties and time pressures in mind.

The researchers tested the model’s performance using critical care patient information from US and European databases to determine outcomes for patients whose actual treatment matched the model’s recommended treatment schedule, and for patients whose actual treatment matched the model’s recommended treatment schedule. compared outcomes in patients who differed from those recommended by For vital signs, test results and risk-related demographic data. Outcome measures were the patient’s survival rates after his 30 and 60 days of sepsis treatment.

“We’ve shown that when real treatments and artificial intelligence match, mortality rates drop. If they don’t agree, mortality rates can be as high as 25%,” says computer science at Ohio State University. and senior author Ping Zhang, Ph.D., assistant professor of engineering and biomedical informatics.

The model was trained and validated on a dataset obtained from a public database called MIMIC-III. The model was tested on various parts of MIMIC-III and a new external dataset from AmsterdamUMCdb. Key measurements in approximately 14,000 sepsis patients included changes in the patient’s vital signs and laboratory test results over time (which serve as indicators of disease severity and type of infection) and antibiotics administered. It included an innovative method designed to compare outcomes in treated and untreated patients. at a specific time.

“We want to model and predict whether antibiotic use at a particular time is beneficial. Yes or no. , applied the clinical trial concept to this model: For every patient who took this drug, we included a clinically similar matched patient who was not taking antibiotics at that time.” , is also a core faculty member of the Translational Data Analytics Laboratory at Ohio State University. “Then we can predict counterfactual outcomes and train counterfactual treatment models to determine whether sepsis treatments are effective.”

Catherine Buck

Sepsis contributes to more than one-third of hospital deaths and is most frequently found in intensive care units and emergency departments. “We often make decisions there without a gold standard due to culture,” says the collaborators. The author, Katherine Buck, M.D., Ph.D., is an assistant professor of emergency medicine in the School of Medicine and director of the geriatric emergency department at Ohio State Wexner Medical Center. “Not all patients who meet criteria for sepsis have evidence of bacterial infection,” she said.

Antibiotics carry risks. It can be harmful to the kidneys, cause allergic reactions, and cause C. difficile, an infection that causes severe diarrhea and inflammation of the colon.

“What the paper begins to say is that the information available to clinicians, sometimes at the forefront and sometimes not, is ‘changing in ways that suggest patients will benefit from antibiotics.’ “A decision-support tool could tell clinicians whether it matches what we already think, or encourage them to ask themselves what they are missing.” I hope that as time goes on, signals will emerge from all the electronic health record data we have, and from there how we use them and how we get them to clinicians. It is important to understand what

These insights, along with the availability of electronic health record data, are key to providing the right kind of data for the model and designing the model to take into account multiple considerations as the healthcare environment changes, Zhang said. says.

“We modeled patient records like language,” he said. “And for machine learning, we always train our models batch-by-batch. We need a model to improve on, and the machine will always find better parameters to fit the model.”

A key metric used to guide how the model reaches its recommendations is the Sequential Organ Failure Assessment (SOFA) score. It is used to routinely assess how the organ systems of ICU patients are performing based on the results of six lab tests. Researchers ran a case study example to demonstrate what an interface developed for a clinical setting might look like, and the model recommended treatment timelines based on changes in personalized patient data. We showed how the SOFA score changes when adjusted.

“Our paper is the first to pursue antibiotic recommendations for sepsis using AI to support clinical decision-making using real-world data,” said Zhang. . “Studies like this require clinical validation. This is Phase 1 of retrospective data analysis, and Phase 2 will require human-AI collaboration for better patient care.”

This work was supported by the National Science Foundation and the National Institutes of Health. Additional co-authors, both from Ohio, include lead author Ruoqi Liu, a PhD student in computer science and engineering, and Jeffrey Caterino, MD, professor and chairman of emergency medicine and director of emergency medical services. )was.

Source link

Leave a Reply

Your email address will not be published. Required fields are marked *