Sepsis is a life-threatening condition that takes the lives of almost a quarter of a million people each year. Able to progress quickly and cause a range of deadly problems, such as tissue damage and organ failure, sepsis is usually caused by the body responding overwhelmingly to a severe infection. Basically, in an attempt to defeat an infection, the body starts attacking itself.
To make matters worse, sepsis often requires critical, timely treatment, though many treatments turn out to be the wrong ones and make things worse. When this happens, patients have started down a “dead-end,” where a specific treatment course will inevitably fail them. The challenge facing clinicians is deciding what treatments to deploy, and when.
Researchers at the Massachusetts Institute of Technology have turned to machine learning and artificial intelligence to see how these tools could aid doctors in the timely, effective application of treatments for sepsis. According to a paper published in Advances in Neural Information Processing Systems, researchers built on previously published work by one of the study’s lead authors, Mehdi Fatemi, to explore how to use AI to predict treatments that could help avoid these dead end scenarios.
As with any AI technology, predictive algorithms need to be “trained”; that is, AI uses datasets from which to draw conclusions. However, researchers noted that a common approach to training AI, reinforcement learning, is too risky and experimental. Reinforcement training is essentially a trial and error learning method, but when it comes to patients in dire need of treatment, guesswork is rather unethical.
Instead, researchers used a small dataset of sepsis patients to help the AI find treatments that didn’t work, which helped the AI show that nearly 12% of treatments administered to sepsis patients actually made things worse.
An additional benefit of this AI tool is it’s timing; according to graduate study Taylor Killian, one of the study authors, “We see that our model is almost eight hours ahead of a doctor’s recognition of a patient’s deterioration. This is powerful because in these really sensitive situations, every minute counts, and being aware of how the patient is evolving, and the risk of administering certain treatment at any given time, is really important.”
The research team also noted that their work could be used in other medical contexts to help optimize treatment solutions, though they emphasize their tool is there to assist, not replace, doctors and clinicians.