A machine learning model a day keeps the sepsis away

AI could be used to optimize sepsis treatment timing via a machine learning model that advises when to administer antibiotics based on various patient data.

Sepsis is a life-threatening complication to an infection in which the body’s immune system overreacts and causes extensive inflammation throughout the entire body. This can promptly lead to tissue damage, organ failure and eventually death if not treated immediately. With conditions like sepsis, patients are running against an extremely hasty clock.

A novel machine learning model developed by scientists from Ohio State University (OH, USA) takes into account these urgent time pressures. Published in Nature Machine Intelligence, the study describes the model’s ability to approximate the most effective time for delivering antibiotic treatment to patients who may go on to develop sepsis. This could help lay the foundations for a new era of precision medicine based on personalized patient data.

Sepsis is a huge problem worldwide, particularly in hospitals and intensive care units. In fact, the World Health Organization states that sepsis is the cause of 1 in 5 deaths globally. Sepsis can be particularly challenging to detect early on as its symptoms such as elevated heart rate and fever, can overlap with numerous other conditions. The main treatment is antibiotics, however they can cause side effects, such as dizziness, nausea, allergic reactions and C. difficile infection. Thus, there is an urgent need to improve the detection and treatment of sepsis.

Now a team at Ohio State University have created a way to improve the effectiveness of sepsis treatment with a machine learning model that estimates optimum treatment timing. The model was validated and trained on a publicly accessible database (MIMIC-III).

The model’s performance was then assessed with information from approximately 14,000 critical-care sepsis patients obtained from a European and a US database. This database stored details regarding symptom severity and infection type, such as patient laboratory test results, vital signs and risk-related demographic data. The outcomes from patient treatments that matched the model’s timing advice were compared to the outcomes from clinically similar patient treatments that differed from the model’s timing advice. Patient outcome was defined as survival 30- and 60-days post sepsis treatment.

By adopting a clinical trial approach when testing the model (through comparing each patient who had antibiotic treatment at the model’s recommended time to a clinically similar patient who had not taken the antibiotic treatment at the recommended time), the team could identify if the model was successful at estimating whether it was favorable to administer antibiotics at a specific time.

“A decision-support tool could tell clinicians if it matches what we’re already thinking or prompt us to ask ourselves what we’re missing. Hopefully, with time, all the electronic health record data we have will reveal signals – and from there it’s a matter of figuring out how to use them and how to get that to clinicians,” noted Katherine Buck, an assistant professor of Emergency Medicine in the College of Medicine at Ohio State.

The results are so far promising. A successful model that takes into account alterations in personalized data and modifies its proposed treatment timeline accordingly could become a key decision-making tool within a clinical setting.

“Our paper is the first to use AI to pursue an antibiotic recommendation for sepsis, using real-world data to help clinical decision making,” concluded Ping Zhang, who leads the AI in Medicine Lab at Ohio State. “Any research like this needs clinical validation – this is phase one for retrospective data analysis, and phase two will involve human-AI collaboration for better patient care.”