
A new study published on 22 October in the journal Nature Medicine presents a computational model called Artificial Intelligence (AI) Clinician to assist with the individualized real-time treatment of sepsis ― also known as blood poisoning ― a severe infection that leads to life-threatening organ dysfunction and the third leading cause of death worldwide (1).
Sepsis is the main cause of in-hospital mortality however the best treatment strategy remains unclear. Apart from the Surviving Sepsis Campaign, a global initiative to improve the treatment of sepsis and reduce the high mortality rates, there is currently no tool to assist clinicians in what is essentially a life-or-death decision-making process and a large body of evidence suggests that “suboptimal decisions lead to poorer outcomes.”
The newly developed technology uses a type of machine learning called reinforcement learning, a category of AI that learns by trial-and-error, to create an optimized set of rules that maximize the desired return ― in this case, the right combination of intravenous fluid and vasopressor doses to maximize the chances of survival over 90 days. A training data is then fed into the neural network to identify patterns and make recommendations based on these patterns. The AI Clinician sifts through each patient’s case and worked out the best strategy for keeping the patient alive based on 48 relevant variables including age, vital signs, and pre-existing conditions.
In the study, researchers from Imperial College London and Harvard-MIT analysed patient records, including data on blood pressure and heart rate, from 130 intensive care units in the US over 15 years period to determine whether the AI system’s recommendations would have been able to improve patient outcomes compared to standard care.
According to the paper, “On average, the AI Clinician recommended lower doses of intravenous fluids and higher doses of vasopressors than the clinicians’ actual treatments.” Around 58 per cent of the predicted vasopressor doses were “very close” to those actually administered by the clinician but the results were not as strong for IV fluid, with only 36 per cent of the recommendations aligned with the clinician’s decision. Not surprisingly, the patient’s chances of survival were highest when the clinician’s treatment matched the recommendations of the AI Clinician.
Importantly, the new tool is not a replacement for medical professionals but could be used alongside existing practices to help doctors decide on the best treatment strategy. The authors write, “Physicians will always need to make subjective clinical judgments about treatment strategies,” but suggest that “computational models can provide additional insight about optimal decisions” to avoid “targeting short-term resuscitation goals” at the expense of longer-term survival.
First, AI Clinical will need to be evaluated in clinical trials using real-time data and decision-making and tested in various healthcare settings but could provide a much-needed approach to guide treatments and improve outcomes. As highlighted by the authors, “a reduction in mortality from sepsis by only a small percentage would represent several tens of thousands of lives saved annually worldwide.”
(1) Komorowski, M. et al. The Artificial Intelligence Clinician learns optimal treatment strategies for sepsis in intensive care. Nature Medicine (2018). DOI: 10.1038/s41591-018-0213-5