The research, published in PLOS One by Junji Shimauchi of the Japanese Red Cross Musashino University and colleagues, took a deliberately different angle from most previous work in this field. Earlier prediction studies typically treated unplanned ICU admission as the outcome of interest, but the authors argue that admission is a location-based measure that can be distorted by institutional realities: bed availability, staffing levels, local admission policies, and even the time of day. A patient may be transferred to the ICU because they need a ventilator, or because there happens to be a free bed and a nervous physician at the bedside. By focusing instead on treatment intensity, the team sought to capture what intensive care actually means physiologically rather than geographically.

The study defined ICU-level interventions with clinical precision. A patient was counted as needing intensive care if, within twenty-four hours of the emergency team activation, clinicians initiated mechanical ventilation, began an infusion of vasopressors or inotropic drugs to support blood pressure and heart function, started renal replacement therapy such as dialysis for failing kidneys, or began polymyxin B hemoperfusion, a blood-purification technique used in severe sepsis. These are treatments that fundamentally alter the level of care a patient receives, regardless of which ward the bed happens to sit on. The second outcome the team modeled was equally stark: death from any cause within fourteen days of the emergency call.

To build their models, the researchers assembled a retrospective cohort of adult patients who experienced Medical Emergency Team activations on general wards at their hospital between May 2021 and December 2025. Crucially, they restricted the predictor variables to information that would realistically be available at the moment the team arrives: vital signs, routine blood tests, and basic clinical characteristics. This constraint matters enormously for real-world usefulness. A prediction model that requires laboratory results that take hours to return, or imaging that a deteriorating patient cannot tolerate, is of little help at the bedside. The final dataset comprised 533 patients, a substantial cohort for a single-center study of this kind.

The numbers reveal how common escalation actually is after an emergency team call. Of the 533 patients, 303, or 56.8 percent, required at least one ICU-level intervention within twenty-four hours, and 147, or 27.6 percent, died within fourteen days. These figures underscore the severity of the population: more than half of all patients whose deterioration triggered an emergency team response went on to need organ support, and more than a quarter did not survive two weeks. The emergency team call, in other words, is not a false alarm in the vast majority of cases, which makes the question of who needs the ICU a matter of routine, high-stakes triage.

Using multivariable logistic regression for the intervention outcome and Cox proportional hazards models for mortality, the team identified a consistent biochemical signature of risk. Lower serum albumin levels, a marker of chronic reserve and nutritional status, and higher blood urea nitrogen, which rises with kidney dysfunction and catabolic stress, were associated with both the need for ICU-level treatment and with fourteen-day mortality. Elevated potassium, often a sign of kidney injury or cellular breakdown, and a high white blood cell count, reflecting infection or systemic inflammation, also predicted both outcomes. Higher heart rate at the time of activation added predictive power specifically for death within fourteen days. The convergence of these markers across both endpoints suggests they capture a common underlying physiology of decompensation, while the divergence of heart rate hints that dying and needing organ support are related but not identical risks.

How well did the models actually perform? For predicting the need for ICU-level intervention, the model achieved a C-index of 0.721, with a 95 percent confidence interval of 0.660 to 0.745, a level the authors characterize as moderate discrimination. For fourteen-day mortality, performance was stronger, with a C-index of 0.784 and a confidence interval of 0.720 to 0.814, indicating good discrimination. The C-index, familiar to clinicians from tools like the Framingham risk score, expresses the probability that a randomly chosen patient who experiences the outcome receives a higher predicted risk than one who does not; a value of 0.5 is no better than a coin flip, while 1.0 is perfect. Calibration analyses, which test whether predicted probabilities match observed event rates, showed modest overfitting but generally good agreement, meaning the models’ risk estimates were broadly trustworthy rather than systematically inflated or deflated.

Perhaps the most clinically forward-looking part of the analysis was the decision curve analysis, a technique that evaluates whether using a model to guide decisions produces net benefit compared with simpler strategies such as treating everyone, treating no one, or relying on clinical judgment alone. Across a meaningful range of threshold probabilities, the models demonstrated net clinical benefit, suggesting that if clinicians used the risk scores to decide which patients to escalate or monitor most intensively, more patients would be correctly triaged than harmed. The team also subjected their models to bootstrap-based internal validation, a resampling technique that estimates how much the apparent performance is inflated by chance fits to the specific dataset, and the results held up reasonably well.

The conceptual contribution of the study may prove as important as its statistical performance. By demonstrating that treatment escalation and short-term mortality are related but distinct dimensions of risk, the authors challenge the assumption that a single severity score can serve both purposes. A patient may be likely to die soon yet not benefit from mechanical ventilation, or may need temporary vasopressor support while having a good prognosis. Separating these endpoints allows clinicians and families to have more nuanced conversations: one about the probability that aggressive organ support will be required, and another about the probability of survival. The treatment-intensity framework also travels better between hospitals, because it does not depend on local bed policies the way admission-based outcomes do.

The authors are careful about the limits of their work. This was a retrospective, single-center study at one Japanese tertiary-care hospital, and both the patient population and local practice patterns may differ elsewhere. The models require external validation in independent cohorts before any clinical implementation, a step that many promising prediction tools never complete. Still, the study arrives at a moment when hospitals worldwide are grappling with strained intensive care capacity and growing interest in algorithmic decision support. If externally validated, a tool that estimates, at the moment of an emergency team call, the likelihood that a patient will need ventilation, vasopressors, or dialysis within a day could help clinicians allocate beds more rationally, frame goals-of-care discussions with families earlier, and ensure that the patients who most need the ICU reach it in time.

Subject of Research: Development and internal validation of clinical prediction models for ICU-level intervention and 14-day mortality following Medical Emergency Team activation

Article Title: Development and internal validation of prediction models for ICU-level intervention and 14-day mortality at Medical Emergency Team activation: A retrospective cohort study

Article References: Shimauchi, J., Nishino, T., Shigeta, K., Mase, H., Yamamoto, T., Yokobori, S., & Ishikawa, M. (2026). Development and internal validation of prediction models for ICU-level intervention and 14-day mortality at Medical Emergency Team activation: A retrospective cohort study. PLOS One, 21(10), e0360511. https://doi.org/10.1371/journal.pone.0360511

Image Credits: AI Generated

DOI: 10.1371/journal.pone.0360511

Keywords: Medical Emergency Team, prediction models, intensive care, ICU-level interventions, 14-day mortality, logistic regression, Cox proportional hazards, decision curve analysis, internal validation, risk stratification, clinical deterioration, retrospective cohort study