How AI can change sepsis detection and care

On May 12, 2026, a quiet regulatory announcement marked what may prove to be a turning point in critical care medicine. Bayesian Health, a company founded by Johns Hopkins AI researcher, Dr. Suchi Saria, became the first continuous AI-based sepsis detection system to receive FDA 501(k) clearance. In practical terms, this means the technology can now move from research environments into routine clinical use.

The system had already received the FDA’s Breakthrough Device Designation, reserved for technologies with the potential to improve treatment for life-threatening conditions. But the clearance represents something more significant: one of the first examples of a real-time clinical AI system meeting the evidentiary standards required for deployment in hospitals at scale.

The condition it targets, sepsis, is one of medicine’s most difficult challenges, not because it is rare of exotic, but because it can hide in plain sight. It occurs when the body’s response to infection spirals out of control, damaging tissues and organs and potentially leading to death within hours. In 2017 alone, an estimated 49 million people worldwide developed sepsis, and around 11 million died from it. Despite modern monitoring systems and highly trained clinicians, sepsis remains notoriously difficult to detect early.

The reason is simple: sepsis rarely presents itself clearly. Its early symptoms such as fever, elevated heart rate, low blood pressure, confusion, overlap with many other conditions, particularly in intensive care units where critically ill patients are already unstable. Blood cultures, still considered the gold standard for confirming infection, can take up to 48 hours. For sepsis patients, that delay can be fatal.

This is where medical AI may offer an advantage.

The promise of machine learning in sepsis detection is not to replace clinicians, but to continuously analyse large numbers of variables simultaneously without fatigue. This way, instead of relying on isolated measurements, AI systems can detect subtle combinations of changes that may signal a patient is deteriorating.

At UC San Diego Health, researchers deployed in emergency departments an algorithm called COMPOSER, publishing their results in npj Digital Medicine in 2024. The system continuously monitors more than 150 variables, including vital signs, laboratory results, medications, demographics, and medical history, calculating in real time the probability that a patient is developing sepsis.

Once the risk crosses a certain threshold, the system alerts healthcare staff through the hospital’s electronic health record system. A physician then reviews the case and decides on the next steps. Importantly, the system also addresses one of the biggest problems in critical care: alarm fatigue. Hospitals already generate enormous numbers of alerts, many of them false positives. Over time, clinicians can become desensitised to warnings. COMPOSER attempts to reduce unnecessary alerts by checking whether the observed data could be better explained by conditions other than sepsis before notifying staff.

Bayesian Health’s FDA-cleared system follows a similar philosophy by integrating into the existing clinical workflows rather than disrupting them. In a Nature Medicine validation study involving more than 764,000 patients across five hospitals, including over 17,000 sepsis cases, the platform demonstrated the ability to integrate into real clinical workflows while continuously monitoring patients in real time. 

Yet detecting sepsis early is only part of the challenge.

Treating sepsis remains highly complex. Questions such as how much fluid to administer, when to use vasopressors, or how aggressively to intervene still lack universally optimal answers. One reason is patient heterogeneity: sepsis does not behave the same way in every individual. Age, immune function, genetics, and underlying conditions all shape how the body responds to infection.

This is where personalised medicine and machine learning begin to converge.

In a 2024 commentary published in Critical Care Explorations, VPH member and critical care physician Finn Catling and colleagues argued that machine learning could help personalise sepsis treatment strategies. Different approaches are already being explored.

Supervised learning systems, trained on labelled clinical data, can be linked to treatment recommendations through human-authored rules. Unsupervised approaches can identify clusters of shared physiological characteristics, but biological meaning and clinical utility are only investigated post hoc and the process can be somewhat speculative. Most ambitiously, reinforcement learning systems, which learn by observing outcomes and receiving signals analogous to reward and penalty, can be trained to optimise for long-term patient outcomes.

But major obstacles remain.

Many AI systems in critical care still rely heavily on a small number of public datasets, such as MIMIC-III, which contains data from a single US hospital collected between 2001 and 2012. Models trained on these datasets may not generalise well to other hospitals, healthcare systems, or low- and middle-income countries where maternal and neonatal sepsis are more prevalent and resources are more limited.

Access to broader and more representative clinical data would help, but healthcare data remains highly fragmented and difficult to share. One possible solution is federated learning, where algorithms are trained across multiple hospitals without patient data leaving its original location.

There are also deeper scientific questions. Machine learning systems are extremely good at finding statistical patterns, but identifying true causal relationships in medicine is much harder. Some researchers therefore advocate combining AI approaches with mechanistic biological models, grounding predictions in known physiology rather than purely statistical correlations.

And even technically strong systems can fail if they are poorly integrated into clinical practice. A recommendation that interrupts workflow, arrives too late, or presents information in an impractical format is unlikely to be used consistently in a busy ICU.

This is why the FDA clearance of Bayesian Health’s system matters beyond sepsis itself. It signals that continuous AI monitoring systems can now meet the regulatory and clinical standards required for real-world deployment.

For years, clinical AI has promised to transform healthcare while struggling to move beyond pilot studies and research papers. This approval suggests that the infrastructure for deployment — regulatory, technical, and clinical — is finally beginning to mature.

Sepsis has resisted medicine’s best efforts for decades, not because the disease is unknowable, but because it evolves faster than conventional systems can respond. AI alone will not solve that problem. But for the first time, tools designed to detect and respond to sepsis continuously, in real clinical settings, are beginning to cross the threshold from experimentation into practice.

The question is how quickly these systems can reach the hospitals and patients that need them most. The hope is that many more similar systems will cross that door.

PS: During the VPH Conference 2024 held in Stuttgart, Germany, Dr. Finn Catling from Imperial College London, then a VPH Student member, received the Best Student Award for his presentation titled “Bayesian inversion enables personalised septic shock treatment guided by noisy arterial pressure waveforms” presenting a system that provides novel cardiovascular insights that could be used to personalise IV fluid, vasopressor and inotrope therapy in septic shock.

Further readings:

• Bayesian Health gets FDA nod for AI sepsis detection tool, MedTechDive, 2026
• Exploring the Potentials of Artificial Intelligence in Sepsis Management in the ICU, Critical Care Research and Practice, 2025
• Machine Learning and Artificial Intelligence in Intensive Care Medicine: Critical Recalibrations from Rule-Based Systems to Frontier Models, Journal of Clinical Medicine, 2025
• Impact of a deep learning sepsis prediction model on quality of care and survival, npj Digital Medicine, 2024
• Can Machine Learning Personalize Cardiovascular Therapy in Sepsis?, Critical Care Explorations, 2024
• Towards personalised treatment in septic shock via Bayesian inversion of a one-dimensional vascular model, Journal of Critical Care, 2024
• Study: AI Surveillance Tool Successfully Helps to Predict Sepsis, Saves Lives, UC San Diego Health, 2024
• Data-driven decision support for individualised cardiovascular resuscitation in sepsis: a scoping review and primer for clinicians, MedRxiv, 2023
• Prospective, multi-site study of patient outcomes after implementation of the TREWS machine learning-based early warning system for sepsis, Nature Medicine, 2022
• Temporal convolutional networks allow early prediction of events in critical care, Journal of the American Medical Informatics Association, 2019