AI Approach Could Cut Sepsis Deaths and Boost A&E Efficiency

King’s College London

By explaining which causes led to a patient's death, the team hope their work will empower hospital managers to make targetted change.

Sepsis Ai thumbnail

Researchers have created a method which could help cut patient deaths by explaining why an AI model 'believes' individuals with sepsis die when processed through A&E.

Designed by researchers in the causal explainability group led by Professor Hana Chockler in the Department of Informatics at King's College London, the method explains why it has made the predictions it has. Utilising real-life data from a hospital, the model identified specific causes it believes contributed to a patient's death, including why they did not receive the correct treatment in time.

Sepsis death in A&E is complicated, with a multitude of factors which may appear unconnected at first leading to it. With the help of this granular, causative data the team hopes the method can empower hospital managers to make targeted changes to improve patient outcomes, prevent deaths and make A&E departments more efficient.

Having a system that [...] isolates causes in the chaos of an emergency room and explains its reasoning to doctors and nurses to build trust with them could make the lives of patients better. This would be as easy as having a dashboard used in A&E."

Dr Nathan Blake

Dr Nathan Blake, a Research Associate in the group, one of the authors of the paper and a former A&E nurse, said "As a nurse in A&E for over five years, I saw a lot of patients go through the department. While every patient journey is unique, after a while you start to see patterns - patterns that if you have a ticking clock like sepsis can mean the difference between life and death.

"Having a system that spots these, isolates causes in the chaos of an emergency room and explains its reasoning to doctors and nurses to build trust with them could make the lives of patients better. This would be as easy as having a dashboard used in A&E."

When a patient arrives at A&E a series of events will occur such as being seen by a nurse to be triaged, having blood taken, or being seen by a doctor. This sequence of events is a patient's 'trace', and due to the complex and fast-moving nature of a hospital emergency room almost all patients are slightly different, even if they present the same problems. All events in a trace, including which member of staff you saw when, can affect the outcome of your time at A&E.

Previously, data scientists have compiled these traces in the lab to train a model that predicts the outcome of patients, e.g. if they should be admitted to the hospital or sent home, through a process called Predictive Process Monitoring. By identifying if a patient needs admission early in their trace, appropriate treatment and a bed are more likely to be found.

Our methods are rooted in a formal and rigorous framework of causality, and they allow a much more accurate and fine-grained analysis of the causes of the delay in the patient's processing. They can also find the most influential causes - those that need to be addressed first to improve outcomes for a stretched hospital."

Dr Tom Yaacov

Dr Tom Yaacov, a Research Associate in the group and the lead author of the paper, said: "Correlation presents in an incomplete picture in a hospital and is of limited use to the managers in charge of improving departments. They lack the rigour to point to specific causes of a patient's deterioration and the ability to explain to a doctor, who may disagree with its conclusions of what factors contributed to a patient's death.

"Our methods are rooted in a formal and rigorous framework of causality, and they allow a much more accurate and fine-grained analysis of the causes of the delay in the patient's processing. They can also find the most influential causes - those that need to be addressed first to improve outcomes for a stretched hospital. Moreover, we can incorporate expert knowledge of medical professionals - something that the existing correlation-based methods cannot do."

The method, presented at the IEEE International Conference on Data Mining, is based on a novel kind of explainability framework that captures the causal dependencies of trace events over time. This allows it to establish and explain a cause-and-effect relationship with what it thinks will be the outcome of those events, the first time this has been done on this scale. These explanations are produced to no negative effect on the computational efficiency of the model, meaning no delays in predicting patient outcomes compared to current approaches.

The work has been funded by the UKRI Causality in Healthcare AI (CHAI) Hub, hence the initial focus on healthcare. However, as the model is agnostic to application, with the same approach applicable across multiple complex cause and effect processes like traffic fines and appeal processes, the group believe that once finalised the model could be used across the healthcare sector and beyond.

/Public Release. This material from the originating organization/author(s) might be of the point-in-time nature, and edited for clarity, style and length. Mirage.News does not take institutional positions or sides, and all views, positions, and conclusions expressed herein are solely those of the author(s).View in full here.