Predicting Sepsis Risk with AI Algorithm
Building and training an AI-based sepsis prediction tool that assesses hundreds of different patient risk factors.
Geisinger Health System, United States
Sepsis is a serious and potentially life-threatening reaction to an infection in the body. It is also called septicemia or blood poisoning and happens when the immune system overacts to an infection and starts to damage tissues and organs. The condition is difficult to spot, and diagnosis isn't clear-cut because many of the signs and symptoms are the same as those of other conditions.
Leveraging Patient Data
Researchers from Geisinger Health System realized that electronic health record (EHR) data contained a huge amount of information that could provide insights into which patients are more likely to die from sepsis, and how best to treat the infection.
To extract these insights Geisinger worked with IBM to build a powerful artificial intelligence predictive model capable of assessing hundreds of different risk factors for every patient.
“For clinicians, making a sepsis diagnosis can be very difficult, as the symptoms overlap with many other common illnesses,” said Dr. Donna Wolk, division director of molecular and microbial diagnostics and development at Geisinger.
“If we can identify patients more quickly and more accurately, we can administer the right treatments early and increase the chances of a positive outcome.”
Earlier diagnosis and rapid treatment can save lives. The mortality risk is raised by 7.6 percent for every hour of delay in sepsis diagnosis.
Creating the Algorithm
To build and train the algorithm the team used open-source technology from IBM Watson while Geisinger provided de-identified files for nearly 11,000 patients diagnosed with sepsis between 2006 and 2016. They broke the data into separate features for each patient covering such details as patient infection types, lifestyle and medical history.
For the next part of the project, the team set themselves the goal of using the data to predict all-cause mortality while patients were in hospital or at home during a 90-day post-discharge period.
After training, testing and fine-tuning the model the team used the model's final version to estimate precision. The results were impressive with the algorithm correctly predicting death for patients in the test data who did die or survival for those whose treatments were successful.
The new algorithm helped researchers identify clinical biomarkers that are associated with higher mortality rates from sepsis. These include age, decreased blood pressure, number of hospital transfers and prior cancer diagnosis. It demonstrated that these were all key factors linked to sepsis deaths.
Geisinger will now continue building on their model by incorporating additional data.
“Thanks to IBM, we are well on our way to breaking new ground in clinical care for sepsis and achieving more positive outcomes for our patients,” added Dr. Wolk.
Next Story »