Artificial intelligence for early sepsis detection in adults: a systematic review of opportunities for clinical decision support and antimicrobial stewardship, and the case for responsible use
This systematic review of 17 studies indicates that while artificial intelligence models demonstrate high accuracy and can detect sepsis significantly earlier than traditional clinical tools, their responsible clinical implementation is currently limited by a high risk of bias, a lack of independent validation, and insufficient evidence regarding their impact on antibiotic prescribing.
Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer
Sepsis is a life-threatening reaction to an infection that can turn a routine illness into a medical emergency within hours. It happens when the body's immune system overreacts, damaging its own organs. Because the early signs—fever, rapid heart rate, confusion—look like many other common ailments, doctors often struggle to spot it quickly. Current tools used at the bedside are imperfect; some catch too many false alarms, while others wait too long until the patient is already in serious trouble. This delay is dangerous because the standard treatment, antibiotics, works best when given immediately. Yet giving them too freely to patients who do not have sepsis fuels a growing crisis of drug-resistant bacteria. The medical community is caught in a difficult balance: waiting too long risks death, but acting too loosely risks creating superbugs that cannot be treated.
In this context, a team of researchers from the University of Hertfordshire set out to examine a potential solution: artificial intelligence. They wanted to know if computer programs could learn from the vast amounts of data hospitals already collect—such as vital signs and blood test results—to spot sepsis earlier than human doctors can. Their goal was not just to see if these programs could predict the illness, but to understand if they could actually help doctors make better decisions about when to start antibiotics and when to hold back. The researchers conducted a systematic review, a rigorous method of gathering and analyzing every relevant study published between 2016 and 2026. They looked at seventeen studies that tested these AI tools in adults across various hospital settings, from intensive care units to emergency rooms and burn centers.
The findings reveal a technology that is remarkably good at its primary job but still unproven in its real-world application. The computer models reviewed were able to identify patients who would develop sepsis several hours before a doctor would typically notice the signs. In the studies where the AI was compared directly to the standard tools doctors use today, the AI consistently performed better, often flagging the risk four to twelve hours earlier. These systems worked across different types of hospitals and patient groups, using nothing more exotic than the routine data already recorded in patient charts. The accuracy of these predictions was high, with the models correctly distinguishing between sick and healthy patients in the vast majority of cases.
However, the researchers found that this high accuracy does not yet guarantee that the technology is ready for widespread use. A closer look at the seventeen studies showed that most of them had significant flaws in their design. The majority were tested only on the same data they were built with, a bit like a student taking a test using the exact same questions they studied from the night before. When the researchers looked for studies that tested these models on completely new groups of patients from different hospitals, the results were much less impressive. The performance of the AI dropped noticeably when it moved to a new environment, suggesting that a program that works perfectly in one hospital might fail in another. Furthermore, very few studies checked if the AI actually helped patients or improved how antibiotics were used.
The review highlights a critical gap between what the technology can do in a lab and what it needs to do in a hospital. While the AI can predict risk, the studies rarely measured whether acting on that prediction led to better outcomes. Only a handful of the reviewed studies looked at whether the AI helped doctors give antibiotics sooner or stop them when they were not needed. In fact, none of the studies reported on whether the technology reduced the overall use of antibiotics or helped doctors choose narrower, more targeted drugs. This is a major concern because if an AI system raises too many false alarms, it could cause doctors to prescribe broad-spectrum antibiotics to everyone, which would worsen the very problem of drug resistance the technology was meant to solve.
The authors conclude that while artificial intelligence offers a genuine and near-term opportunity to save lives by catching sepsis earlier, the path to responsible use is not yet clear. The technology has proven it can see the warning signs before humans do, but it has not yet proven it can be trusted to guide treatment decisions in diverse hospitals. To move forward, the researchers argue that future studies must test these models in the specific hospitals where they will be used, rather than just in the labs where they were created. They also emphasize that doctors and hospitals need to measure whether these tools actually change prescribing habits for the better, ensuring that the rush to treat sepsis does not come at the cost of creating harder-to-treat infections. Until these steps are taken, the promise of AI in sepsis care remains a powerful tool waiting for the right conditions to be fully realized.
Drowning in papers in your field?
Get daily digests of the most novel papers matching your research keywords — with technical summaries, in your language.