The piece of research that we have just published in Nature Medicine demonstrates how a close collaboration between clinicians and data scientists can disrupt the current paradigm for generating medical evidence and lead to the development of meaningful clinical tools with the potential to improve patient outcomes.
Sepsis is the main cause of mortality in hospitals and the most expensive condition treated in hospitals. Estimates of sepsis deaths reach 6 million per year worldwide, including about 260,000 in the US and 44,000 in the UK.
Besides early control of the source of infection (with antibiotics and surgery when necessary), a key aspect of sepsis treatment relies on the correction of hypovolaemia (low circulating blood volume) and vasoplegia (loose arteries) via the administration of intravenous fluids and vasopressors. However, timing and titration of these medications is extremely tricky in clinical practice, with consistent evidence that 1) patients frequently receive the wrong dose, and that 2) this is associated with poorer outcomes.
To address this sequential decision-making challenge, we modelled the dynamics of about 96,000 patients with sepsis coming from over 130 intensive care units in the US over a 15-year period. Then, we deployed reinforcement learning algorithms to estimate an optimal treatment strategy. In a cohort independent from the training data, and using state of the art evaluation methods, we were able to show that the AI system made, on average, more reliable treatment decisions than human doctors.
This publication is only the start of a long and complex journey to bring back the algorithm to the bedside. We hope to gain further evidence of the value of our approach by deploying it in the clinical environment, and we are exploring various avenues to achieve this. Our vision is that this new tool will be used alongside medical professionals, to support (and not replace) doctors decide the best treatment strategy by reducing the uncertainty around sepsis treatment. Importantly, the approach is agnostic to the question asked, and could be applied to many clinical environments and clinical problems, which we also hope to develop in the future.
Summary by Matthieu Komorowski:
MD with full board certification in anaesthesiology and critical care in both France and the UK, Matthieu was formerly a research fellow at the European Space Agency. He currently pursues a PhD at Imperial College and a research fellowship in intensive care at Charing Cross Hospital in London. A visiting scholar at the Laboratory for Computational Physiology at MIT, he collaborates with the MIT Critical Data group (Professor Leo Celi) on numerous projects involving secondary analysis of healthcare records. His research brings together his expertise in machine learning and critical care to generate new medical evidence and build decision support systems, with a particular interest in sepsis.