Hospital prediction systems can identify deteriorating patients, but their value depends on the response they trigger. Recent studies reveal both measurable clinical benefits and the continuing burden of false alerts.
An early warning only helps if it arrives when something useful can still be done. In acute care, an algorithm may identify a patient at increased risk, yet add little if the clinical team has already recognised the problem. Conversely, a modest signal can matter if it brings the right person to the bedside sooner.
This makes sepsis and hospital deterioration demanding tests for artificial intelligence. The technology must operate amid changing observations, incomplete information and competing priorities. Its output becomes one more claim on staff attention, alongside direct examination, existing escalation systems and the concerns of patients and relatives.
The research is becoming more informative. Prospective implementation studies and randomised trials have begun to examine mortality and escalation of care, while newer external validations show why apparently strong predictive performance can coexist with a substantial alert burden. For European hospitals, international evidence offers concrete lessons about what to test before introducing another warning into an already busy clinical environment.
A familiar model has changed since its first scrutiny
A 2021 external validation of the Epic Sepsis Model became a prominent cautionary example. Researchers examined 38,455 hospitalisations involving 27,697 patients at Michigan Medicine. The model’s area under the receiver operating characteristic curve was 0.63, indicating substantially weaker discrimination than its earlier reported performance.
That finding concerns a historical version and setting. It should not be used as though it measured every subsequent product version. In February 2026, a multicentre prospective validation examined version two across 227,091 inpatient encounters at four US health systems. Encounter level discrimination ranged from 0.82 to 0.92, but positive predictive values remained between 13% and 26% at the evaluated operating thresholds.
The improvement and the limitation can both be true. A model can rank patients’ risk substantially better while still flagging many encounters that do not meet the outcome definition. Positive predictive value describes how often a positive classification corresponds to that outcome, and is affected by its prevalence as well as the chosen threshold.
For a hospital, the practical question is how many reviews that threshold creates and what staff learn from them. An impressive overall ranking statistic does not specify an affordable or clinically appropriate volume of interruptions.
Earlier recognition is a more demanding comparison
The updated validation also compared model timing with clinical recognition, represented by orders for antibiotics, lactate or cultures. Median lead times among true positive cases varied between institutions. This is a more useful comparison than asking whether a prediction precedes a later coded diagnosis, although clinical orders remain an imperfect proxy for when someone first became concerned.
The distinction changes what counts as useful novelty. A system that reacts to information generated by an ongoing sepsis work up may be accurately recognising an existing clinical response. A system that brings previously overlooked deterioration to attention offers a different contribution.
European evaluations should therefore record both when a warning appeared and what was already happening. They should also follow cases that did not trigger a warning. A service cannot understand an alert’s clinical value solely by studying the patients it successfully identifies.
This requires collaboration between clinicians and developers on the definition of an actionable event. Threshold selection should reflect the intended response, whether an immediate review or closer observation, and the ability of the local team to deliver it. Changing that response changes the intervention being evaluated.
TREWS links response with outcomes but cannot isolate causation
The TREWS study, published in Nature Medicine in 2022, monitored 590,736 patients across five hospitals. Its outcome analysis focused on 6,877 patients with sepsis identified by an alert before antibiotics were initiated. Patients whose alert was confirmed by a provider within three hours had an adjusted 3.3 percentage point lower in hospital mortality than those without confirmation within that period. The adjusted relative reduction was 18.7%. TREWS prospective cohort study
These are associations within an implementation cohort, not results from randomly assigning early confirmation. Differences in recognition, clinical presentation or the circumstances allowing staff to respond could influence the comparison despite adjustment.
The study nevertheless makes a useful operational contribution: the clinical response is visible in the analysis. It draws attention to the interval between a warning and a professional judgement, rather than treating an alert generated in the record as a completed intervention.
A local service could examine that interval to identify avoidable delays or unclear responsibility. The objective would be timely assessment of the patient, rather than achieving a high rate of electronic acknowledgement without understanding whether care changed.
COMPOSER tests a whole emergency department workflow
At UC San Diego Health, COMPOSER was evaluated in two emergency departments using a prospective before and after design. The 2024 report found an observed sepsis mortality rate of 9.49% during the five month intervention period, compared with a modelled counterfactual of 11.39%. That represents an estimated absolute difference of 1.9 percentage points and a relative difference of approximately 17%. COMPOSER clinical implementation
The comparison depends on a statistical estimate of what would have happened without deployment, rather than a concurrent randomised control group. Mortality improvement was statistically significant at the safety net hospital, but not at the other clinical site.
The intervention also involved preparation, education and feedback for nursing staff. Alerts were linked to a defined workflow, with options reflecting whether infection was suspected, a doctor would be notified or treatment was already underway.
That combination is relevant to European adoption. Buying the model does not automatically reproduce the implementation. Hospitals need to understand the response process and the conditions under which the reported benefit occurred. Equally, evaluation should not attribute the entire change to software if training or other improvements contribute.
Nursing observations can become a shared signal
CONCERN takes a different approach to identifying deterioration. Its machine learning system uses patterns in nursing surveillance documentation and displays risk within the clinical record. A pragmatic cluster randomised trial assigned 74 clinical units across two US health systems to the intervention or usual care. The published 2025 analysis included 60,893 hospital encounters.
The adjusted hazard ratio for in hospital death was 0.64, with a 95% confidence interval of 0.53–0.78. The intervention also increased the instantaneous risk of unanticipated intensive care transfer, with an adjusted hazard ratio of 1.25. These are relative time to event measures, not absolute percentage point changes.
The combination matters. More intensive care transfers need not mean that an early warning system has failed; they can accompany earlier escalation for patients who need additional support. However, the result does not establish that every additional transfer was necessary or explain every mechanism behind the mortality difference.
It also challenges the idea that AI’s contribution must come from discovering information clinicians cannot perceive. Here, part of the opportunity is making patterns associated with nursing concern more visible to the wider team. Transfer to another system would require understanding how its nurses document and communicate.
A mortality trial offers a different kind of proof
In Taiwan, a pragmatic randomised trial tested alerts generated from AI analysis of electrocardiograms. It involved 39 physicians and 15,965 patients. Ninety day all cause mortality was 3.6% with the intervention and 4.3% under control conditions, corresponding to a hazard ratio of 0.83 with a 95% confidence interval of 0.70–0.99.
This was an alert intervention for elevated mortality risk, not a sepsis diagnostic system. The finding provides randomised evidence that a prediction linked to clinical action can improve an important outcome in a defined setting. It does not establish that all risk scores or alert formats will do so.
Across these studies, the product being tested is increasingly a clinical process: information, communication, review and the capacity to act. For hospitals choosing their next investment, the decisive evidence will show who receives earlier useful care, what additional work is created and whether the resulting balance improves outcomes. An early warning becomes valuable at the point where that chain is completed.
