You are currently viewing New AI predictive algorithm increases sepsis prognosis timeliness and accuracy

New AI predictive algorithm increases sepsis prognosis timeliness and accuracy


Sepsis affects around 30 million people worldwide each year, resulting in an estimated six million deaths. Sepsis is a life-threatening condition caused by the body’s overreaction to an infection.

Because each hour of delayed treatment increases the risk of death by four to eight percent, accurate and prompt sepsis forecasts are critical for reducing morbidity and mortality. To that goal, a number of health-care institutions have used predictive analytics to identify patients with sepsis using data from electronic medical records (EMRs).

Data scientists, physicians, and engineers from McMaster University and St. Joseph’s Healthcare Hamilton collaborated to develop an Artificial Intelligence (AI) predictive system that dramatically increases the speed and accuracy of data-driven sepsis forecasts.

“Sepsis can be predicted very accurately and very early using AI and clinical data,” said Manaf Zargoush, study co-author and assistant professor of health policy and management at McMaster’s DeGroote School of Business. “How much historical data these algorithms need to make accurate predictions and how far ahead they can predict sepsis accurately are the key questions for clinicians and data scientists.”

Some systems combine EMR data with disease scoring methods to calculate sepsis risk scores in clinical care settings, basically operating as digital, automated evaluation tools. Predictive analytics, such as AI algorithms, are used in more advanced systems to go beyond risk assessment and diagnose sepsis itself.

Researchers developed the Bidirectional Long Short-Term Memory method using AI predictive analytics (BiLSTM). Administrative factors (e.g., length of ICU stay, hours between hospital and ICU admission, etc.), vital signs (e.g., heart rate and pulse oximetry, etc.), demographics (e.g., age and gender), and laboratory tests are all examined (e.g., serum glucose, creatinine, platelet count, etc.). The BiLSTM is a more advanced type of machine learning called deep learning that uses neural networks to boost its predictive potential when compared to other algorithms.

The BiLSTM was compared to six other machine learning algorithms in the study, and it was shown to be more accurate than the others. Improving accuracy by eliminating false positives is critical for a successful algorithm, as these errors not only waste medical resources but also undermine clinicians’ trust in the algorithm.

Surprisingly, the study discovered that algorithms that focus more heavily on a patient’s current datapoints, rather than searching back further to include as many datapoints as possible, can improve forecast accuracy.

It’s understandable that therapists would want to fill the algorithm with as many data points as possible over a long period of time, according to the researchers. Their findings show, however, that when the goal of prediction is to be precise and timely in sepsis forecasts, clinicians with extended prediction horizons should depend more on the patient’s fewer but more recent clinical data.

“In late November, St. Joe’s Healthcare Hamilton will launch a cognitive computing pilot project to see how AI can be used to assist forecast sepsis in real patients and in real time,” said Dan Perri, research co-author, physician, and chief information officer. He is also a medical associate professor at McMaster University.

Understanding the breadth and scope of data that enables sepsis prediction is important for any organization looking at using AI to save lives from severe infections.”

Dan Perri, study co-author, physician, and chief information officer at St. Joseph’s Healthcare Hamilton

“Lessons learned from sepsis models are translated into better machine learning tools that lead to effective early response for some of the sickest patients while also avoiding superfluous alarms that could lead to health-care provider fatigue.”


McMaster University

Journal reference:

Zargoush, M., et al. (2021) The impact of recency and adequacy of historical information on sepsis predictions using machine learning