An AI Tool at Johns Hopkins Just Predicted Which ICU Patients Would Deteriorate 12 Hours Before Doctors Noticed

If you walk through any busy intensive care unit on a weeknight, you’ll see a system built to detect deterioration as it occurs: rows of beds divided by curtains, monitors cycling through numbers, nurses moving between patients carrying IV bags and tablets. Sepsis presents a problem because it doesn’t always work with that design. For hours, the infection that sets off the body’s cascading immune response can appear to be many different things. fever. bewilderment. a heart rate that is marginally higher. It is simple to identify the underlying illness that caused the patient’s admission. An hour or two has gone by by the time the pattern resolves into something clearly septic. Furthermore, every hour counts when it comes to sepsis; studies have repeatedly shown that a 3.6–9.9% increase in mortality is linked to every hour that antibiotic treatment is postponed.

Suchi Saria, the founding research director of the Malone Center for Engineering in Healthcare at Johns Hopkins, created the Targeted Real-Time Early Warning System specifically to address that issue. From the time a patient enters the hospital until they are discharged, the system continuously scans their electronic medical records, extracting information from clinical notes, lab results, vital signs, and medication histories. It does not create a risk score at the time of admission. It never stops running. The system generates an alert for the clinical team and updates its assessment in real time if the data pattern changes.

Johns Hopkins ICU AI & Sepsis Prediction: Key Facts

The study’s findings, which were published in 2022 in Nature Medicine and Nature Digital Medicine, are remarkable enough to demand close examination. TREWScore was accurate in 82% of sepsis cases and correctly flagged patients at a rate nearly 40% higher than comparable approaches across 590,000 patients treated at five hospitals with over 4,000 clinicians using the tool. The accuracy of earlier electronic sepsis detection tools ranged from 2 to 5 percent. The AI identified deterioration an average of almost six hours ahead of traditional clinical recognition in the worst cases, where a missed hour is not just a setback but a point of no return. The risk of death was 20% lower for patients whose sepsis was identified by TREWScore.

“It is the first instance where AI is implemented at the bedside, used by thousands of providers, and where we’re seeing lives saved,” Saria said in terms that are worth pondering. That is not an assertion regarding a successful trial conducted in a controlled study setting. It describes a program that is currently operating in real hospitals with real clinical settings and quantifiable results for real patients.

One of the reasons this is so important is the false alarm issue. ICUs already have a ton of alerts. Researchers refer to this mental tiredness as “alarm fatigue” for nurses and doctors. It is caused by a continuous barrage of beeping monitors, warning machines, and flagging systems, the majority of which are either erroneous or not immediately actionable. Clinicians lost faith in earlier AI sepsis detection tools because they were producing false alarms so frequently. In order to specifically address this, TREWScore was created with a transparency feature that explains to clinicians why the algorithm is concerned and which particular data points are raising the alarm. Beyond merely satiating curiosity, this transparency enables clinicians to apply their own judgment to the output instead of just reacting to or disregarding a binary alarm.

The extent to which that transparency feature actually alters clinical behavior on a large scale is still unknown. While 95% of clinicians responded within 24 hours to TREWScore alerts during the pilot phase, only roughly 27% of alerts led to direct clinical action, according to a study. Physicians and nurses must make decisions about how much weight to give machine-generated recommendations, how to integrate them with their own assessments, and how much more chart review is necessary before taking action. This delay between receiving an alert and acting upon it reflects actual conflicts in clinical workflow. In-hospital mortality decreased by 3.3 absolute percentage points when clinicians responded to TREWS alerts within three hours, according to a different analysis. In other words, the algorithm’s usefulness is largely dependent on how its outputs are actually used by the people who receive them.

In order to lower the barrier to adoption at other hospitals, Bayesian Health, a company that was spun out of Johns Hopkins to oversee TREWScore’s clinical deployment, has worked to integrate the system with Epic and Cerner, the two biggest electronic health record platforms in the US. In order to identify patients at risk for conditions other than sepsis, such as pressure injuries, acute respiratory failure, cardiac arrest, and ICU delirium, the team has also modified the underlying architecture. Early in 2023, a different Johns Hopkins project was announced that was specifically focused on predicting ICU delirium, a condition that affects a large percentage of critically ill patients, greatly raises the risk of death, and is infamously difficult to predict using traditional clinical observation alone.
As this work progresses, it seems as though the core framework of ICU monitoring is evolving in ways that will take years to fully take in. The conventional approach places humans at the center of pattern recognition, with doctors performing rounds and synthesizing what they see, and skilled nurses reading a patient’s condition throughout a shift. The volume and complexity of data generated by a modern ICU patient exceeds what any attentive human can reliably track continuously, which is why that model is under pressure rather than because it is flawed. The nurse at the patient’s bedside is not replaced by TREWScore. It keeps an eye on everything at once, never grows weary, and reveals the pattern before it becomes apparent. It is still up to the clinician to decide whether to take action based on what the algorithm reveals. However, they are now, for the most part, given the opportunity to arrive on time.