Sepsis: an AI early warning linked to fewer deaths in hospital

Why does sepsis call for such fast action?

Sepsis, often called blood poisoning, is a severe reaction of the body to an infection that can damage organs. The authors recall that recognising and treating sepsis early improves patients' outcomes.

But it is hard to spot among dozens of records, tests and vital signs to follow. Early warning systems based on machine learning could shorten the time to recognition, yet few have been evaluated in real clinical settings.

What did the researchers do?

They studied an alert system already in use, called TREWS (Targeted Real-time Early Warning System). It continuously analyses data in the medical record to flag patients who might be developing sepsis. The study is prospective: results were followed as they happened, in five hospitals.

In total, 590,736 patients were monitored by TREWS. The authors focused on 6,877 patients with sepsis who were identified by the alert before antibiotics were started. They compared those whose alert was confirmed by a clinician within 3 hours with those whose alert was not, taking into account the patient's condition on arrival and severity.

The study period was long enough to follow what happened to these patients until they left hospital, including organ failure and how long they stayed.

What were the results?

Among patients whose alert was confirmed within 3 hours, in-hospital mortality was 3.3 percentage points lower (confidence interval 1.7 to 5.1), a relative reduction of 18.7%. Organ failure and length of stay were also reduced.

In patients who were also flagged as high risk, the improvement in mortality was larger: 4.5 percentage points (interval 0.8 to 8.3). This suggests that, at the time of the alert, it is possible to identify those who would benefit most from quick treatment.

What could this mean for patients?

A system that flags possible sepsis early can help teams react faster, especially if a clinician reviews the alert quickly. In this study, it is the human response to the alert, within 3 hours, that is linked to better results.

So the AI does not act alone: it draws attention, and it is the clinicians who confirm, decide and treat.

What are the limits of this study?

The study compares groups of patients according to whether the alert was confirmed quickly or not, without random assignment. The patients are therefore not strictly comparable, even though the authors adjust the results: other factors, such as the attention given to a patient, may play a part. The authors speak of an association and a potential for improvement.

It covers five hospitals in one country. A randomised trial would be needed to establish with certainty that the alert, and not other differences between patients, improves survival.