The AI That Catches Sepsis Hours Before Doctors Do

Sepsis is one of the deadliest and most elusive killers in modern medicine. It happens when an infection triggers a chain reaction through the body, causing inflammation, blood clots, leaking blood vessels, and eventually organ failure. In the United States alone about 1.7 million adults develop sepsis every year and more than 250,000 of them die. The faster it is caught, the better a patient's chances of survival. An AI system from Johns Hopkins University, tested on hundreds of thousands of real hospital patients, is proving that machines can see the warning signs hours before human clinicians do, and that those hours translate directly into lives saved.

The system is called TREWS, the Targeted Real-Time Early Warning System. It was developed by Suchi Saria, a Johns Hopkins researcher who lost her own nephew to sepsis when he was a young adult. The machine learning system combines a patient's medical history with current symptoms and lab results, continuously scanning records and clinical notes to identify patients at risk and suggest treatment protocols such as starting antibiotics. In the landmark study published in Nature Medicine and Nature Digital Medicine, patients whose TREWS alerts were confirmed by clinicians were 20 percent less likely to die of sepsis.

Why Sepsis Is So Easy to Miss

Sepsis is a time bomb. Its early symptoms, fever, confusion, rapid heart rate, are common in dozens of other conditions. A patient admitted for a routine infection can slide into septic shock within hours, and by the time the signs are unmistakable, the window for effective treatment may be nearly closed. Under the current standard of care, sepsis kills roughly 30 percent of the people who develop it, largely because detection comes too late.

Hospitals have tried electronic alert tools before, and the results were dismal. Earlier rule based systems caught less than half of eventual sepsis cases, and when they did fire, they were accurate only 2 to 5 percent of the time. Clinicians learned to ignore them. Constant false alarms are worse than no alarms at all, because they erode trust exactly when attention matters most.

A Real World Trial at Scale

What separates TREWS from earlier efforts is the scale and rigor of its validation. Over two years, more than 4,000 clinicians at five hospitals used the system while treating 590,000 patients. The team also retrospectively reviewed 173,931 previous patient cases to benchmark performance. The results, published in 2022 in Nature Medicine and Nature Digital Medicine, represented the first instance of an AI tool implemented at the bedside, used by thousands of providers, with measurable lives saved.

The numbers are striking. In the most severe sepsis cases, where a single hour of delay can mean death, TREWS detected the condition an average of nearly six hours earlier than traditional methods. The system tracks patients from the moment they arrive at the hospital through discharge, so critical information is not lost when shifts change or a patient moves between departments, two of the most common moments for warning signs to slip through the cracks.

Trustworthy by Design

One reason earlier sepsis tools failed is that they acted as black boxes. A flashing alert with no explanation invites dismissal. TREWS takes the opposite approach: it shows clinicians why it is making a specific recommendation, surfacing the evidence in the patient's record that triggered the warning. The doctor remains the decision maker. The AI is a tireless assistant that reads everything, all the time, and explains its reasoning.

That transparency is central to why adoption worked at scale. When an alert says which lab values, symptoms, and history point to sepsis risk, clinicians can verify the logic in seconds and act with confidence. It is a model for how clinical AI should be built: not to replace judgment, but to feed it better information faster.

From Research to Wards Everywhere

To carry the technology beyond the study hospitals, Johns Hopkins spun off Bayesian Health, which led and managed the deployment across all testing sites. The team partnered with Epic and Cerner, the two largest electronic health record providers in the United States, so that any hospital already running those systems can implement the tool without rebuilding its infrastructure. Johns Hopkins Technology Ventures reported in 2025 that the platform is reducing sepsis mortality by roughly 18 percent across dozens of hospitals, a figure consistent with the original study findings at scale.

The approach is also spreading beyond sepsis. The same underlying platform has been adapted to identify patients at risk for pressure injuries, commonly known as bed sores, and other complications that develop quietly during hospital stays. Each new application follows the same formula: continuous monitoring, explainable alerts, and rigorous measurement of real patient outcomes.

The Human Story Behind the Code

Suchi Saria's motivation is deeply personal. Her nephew died of sepsis as a young adult. By the time doctors detected what was happening, he was already in septic shock. Sepsis develops very quickly, she has said, and that speed is exactly what the system is built to beat. The tragedy of sepsis is that it is highly treatable when caught early. Antibiotics and fluids work. What fails is the clock.

TREWS is what happens when that clock gets a machine on the other side of it. Not a distant promise of AI transforming medicine, but a deployed system, running in production wards, credited in peer reviewed literature with saving lives. Sepsis will never be an easy diagnosis. But for hundreds of thousands of patients, the window for survival just got several hours wider.