The Artificial Intelligence Clinician learns optimal treatment strategies for sepsis in intensive care. - Komorowski et al., Nature Medicine, Nov 2018

Original article:

One-sentence summary

In this paper, a research group from Imperial trained an AI algorithm by reinforcement learning to recommend doses of IV fluids and vasopressors to be given in sepsis, and showed that clinical outcomes are better when the AI's recommendations are followed

What did they do?

They used patient data from a 72-hour period around the suspected onset of sepsis, which included 48 variables (demographics, vital signs, lab values, etc) coded as discrete time series (with 4 hour steps). The AI was trained using a Markov Decision Process, with the algorithm being rewarded for survival and punished when a patient died.

The algorithm on average recommended lower doses of IV fluids and high doses of vasopressors. Early use of low-dose vasopressors and high use of IV fluids have both been associated with poorer outcomes in the past. The decisions recommended by the AI were shown to correlate with improved outcomes.

What does this mean?

This algorithm will need prospective clinical validation. If the model performs well, it could be used in clinical settings to guide the amount of IV fluid and  vasopressor given to improve outcomes in sepsis.