AI Case Study
St. Michael’s Hospital predicts onset of sepsis among ICU patients with 90% accuracy using learning evolutionary algorithm framework
Researchers at MIT CSAIL harnessed the power of LEAF(Learning Evolutionary Algorithm Framework) developed by Sentient Labs to use physiologic data to predict onset of sepsis among patients in ICU with 90% accuracy. The evolutionary algorithm analyses arterial blood pressure and heart rates measured minute by minute to predict sepsis onset approximately 30 minutes before.
Healthcare Providers And Services
"Over a one-year period, Sentient Labs worked with MIT’s Computer Science and Artificial Intelligence Laboratory team (CSAIL) to apply evolutionary algorithms on a massive scale to evolve classification rule-sets with high accuracy and acceptable false-positives over unseen data.
Applying these rule-sets to more than 6,000 patient records (representing 4TB of data), St. Michael’s Hospital at the University of Toronto was able to predict thirty minutes ahead of time, with greater than 90 percent accuracy, whether a patient would get sepsis. This provided caregivers valuable time to proactively treat the infection and save patients’ lives.
MIT CSAIL used Sentient’s unique evolutionary algorithm, mapped
across tens of thousands of nodes, that scales to vast resources and addresses highly complex problems to interpret and accelerate analysis of per second heart rate data and arterial blood pressure data to predict onset of sepsis with over 90 percent accuracy and only 12 percent false positives"
"Sepsis affects a million patients a year in the US resulting in deaths of 30-50%. It is also the most expensive condition treated
in U.S. hospitals. Even though there are signs pointing to the onset it was too complex to analyze until now."
The company claims:
* 90% accuracy in predicting Sepsis onset among 6000 patient records studied
* Only 12% false positives
"Inspired by biological evolution, LEAF (learning evolutionary algorithm framework) can evolve any data to solve specific problems. One example of this is evolved neural networks (ENNs)"
"More than 6,000 patient records (representing 4TB of data)"
" physiologic data sets (heart rate/arterial blood pressure)"