New AI Risk Model Predicts Death in Emergency Patients With Striking Accuracy

Emergency departments around the world are under relentless strain, and one of the hardest questions clinicians face every shift is deceptively simple: which patients are most likely to die? A team of New Zealand researchers has now built a data-driven answer. In a study published in the Journal of Medical Systems, scientists from Auckland University of Technology and the University of Waikato introduce IMPACT, the Integrated Model for Patient Acuity, Care, and Triage Efficiency, a statistical framework that crunches nearly six hundred thousand emergency department records to generate individual mortality risk scores across three clinical timeframes.

The scale of the underlying dataset is what sets the work apart. The researchers analysed 593,673 de-identified emergency presentations to Waikato Hospital, a tertiary facility serving a large and ethnically diverse region of New Zealand, between January 2016 and August 2024. After cleaning an initial pool of 724,830 records, the team retained variables spanning patient demographics, triage scores, clinical complexity measures, waiting times, and lengths of stay, together with mortality outcomes recorded within the department, within ten days of discharge, and within twenty-eight days of discharge.

Methodologically, IMPACT takes a deliberately transparent two-step approach rather than relying on opaque black-box artificial intelligence. First, a Generalised Linear Model identified which predictors significantly influence mortality risk and in what direction. Second, the team applied regularised logistic regression techniques, including LASSO, Ridge, and ElasticNet, to sharpen predictive accuracy while taming multicollinearity among correlated variables. The study followed the internationally recognised TRIPOD reporting guidelines, and the final risk score was built on the ElasticNet model, which combines feature selection with coefficient shrinkage to produce stable, interpretable estimates.

The performance figures are remarkable. On a held-out test set comprising twenty percent of the data, IMPACT achieved an area under the receiver operating characteristic curve of 0.9532 for deaths occurring in the emergency department itself, 0.9183 for mortality within ten days, and 0.8999 for mortality within twenty-eight days. Accuracy ranged from 0.90 for immediate department deaths down to 0.83 for the twenty-eight-day horizon, with balanced precision and recall across classes despite the very low baseline event rates of less than one percent.

Among the strongest and most consistent predictors was advanced age. Deaths concentrated overwhelmingly in patients aged seventy to ninety, reflecting frailty, comorbidities, and reduced physiological reserves. Higher clinical complexity scores, normalised across two different national coding systems using Z-score transformation, also raised mortality odds in every timeframe, as did longer stays in the department. Male sex carried a modestly elevated risk, while higher triage scores, which indicate lower clinical urgency, showed a strong protective effect, validating the ability of triage systems to correctly sort the sickest patients first.

One finding will surprise many readers: longer waiting times were associated with lower mortality odds. The researchers are careful to stress this is an associational pattern, not a causal one. The explanation lies in triage-induced prioritisation. The most acutely unwell patients are seen almost immediately, so a short wait is actually a marker of severe illness and high intrinsic risk. Conversely, patients who wait longer tend to be those who were less critical to begin with. Waiting time, in other words, functions as an operational correlate of acuity rather than an independent protective factor, a nuance the authors say challenges simplistic uses of wait-time metrics as measures of care quality.

The data also exposed persistent ethnic health inequities. Māori patients constituted 29.1 percent of presentations and showed moderately elevated mortality odds, particularly over the longer ten- and twenty-eight-day windows, while Pacifica patients also displayed increased odds. Both groups arrived with the highest average clinical complexity scores among deceased patients, suggesting they may reach hospital with more advanced disease. These results align with earlier national research documenting higher mortality and re-presentation rates among Māori, and they give the IMPACT framework a concrete equity-monitoring role: aggregated scores could flag disproportionate risks and inform culturally responsive care pathways and targeted resource allocation.

Temporal patterns in the dataset added further context. Patient arrivals climbed steadily from 2016 to 2019, dropped sharply at the start of the COVID-19 pandemic, then recovered and stabilised at a slightly lower level from late 2020 onwards, with seasonal winter peaks. Meanwhile, both average length of stay and waiting time trended upward across the study period, with pronounced spikes during 2020 and 2021 and a steep rise toward the end of 2024, signals the authors interpret as growing systemic pressure, whether from rising case complexity or operational bottlenecks.

In its current form, the researchers emphasise, IMPACT is a retrospective risk-stratification and service-evaluation tool, not a bedside decision-support system. A real-time version would need to exclude length of stay, which is only known at discharge, and restrict itself to information available at the moment of triage. Formal calibration assessment has not yet been conducted, and the team acknowledges other limitations: missing waiting and stay times were imputed with zero, missing triage scores were assigned the lowest urgency category, cause-of-death data were unavailable, and the single-hospital design may limit generalisability. Socioeconomic status, comorbidity detail, and staffing levels were absent from the dataset and may act as unmeasured confounders.

Even so, the trajectory is clear. The authors envision IMPACT embedded within electronic health records, continuously scanning incoming patients and firing dynamic clinical alerts when calculated risk crosses thresholds, potentially extending into inpatient wards, surgical triage, and chronic disease management. As emergency systems worldwide grapple with ageing populations, rising demand, and widening inequities, a transparent scoring model that quantifies risk, exposes systemic bottlenecks, and holds up a mirror to health disparities may prove one of the most quietly powerful tools emergency medicine has produced.

Subject of Research: A predictive risk-scoring model for mortality and triage efficiency in emergency departments

Article Title: IMPACT: Integrated Model For Patient Acuity, Care, And Triage Efficiency In Emergency Department

Article References: Rasouli Panah, H., Ijadi Maghsoodi, A., Madanian, S., & Yu, J. (2026). IMPACT: Integrated Model For Patient Acuity, Care, And Triage Efficiency In Emergency Department. Journal of Medical Systems, 50(1), Article 135. https://doi.org/10.1007/s10916-026-02420-2

Image Credits: AI Generated

DOI: 10.1007/s10916-026-02420-2

Keywords: emergency department, mortality prediction, triage, risk score, logistic regression, ElasticNet, health inequities, Māori health, length of stay, waiting time, New Zealand, predictive modelling