An alarm needs an answer
Epic's deterioration score has FDA clearance. It belongs only on wards where a staffed team answers its top alerts, the medical staff sets its thresholds and the hospital counts whom it helps and misses.
Listen to this special topic
read by an AI voiceThe Food and Drug Administration has cleared the deterioration score that Epic builds into its health record. Hospitals have used the index for years; on Sept. 18 the agency found its second version substantially equivalent to a competing score, eCART. The software reads 195 items in a patient's record and scores adults on hospital wards or under observation for the risk of needing a rapid response team, a code or a move to intensive care within a day, or of dying within three days. Each hospital sets its own alert thresholds, and the clearance summary leaves that choice to trained professionals at each organization, adding that the score is not to be the sole basis for decisions that may cause death or an irreversible decline in health.
That leaves the two decisions that matter most to each hospital: where to draw the line, and who answers when a patient crosses it. Where such scores have been tied to fewer deaths, the gains came with people assigned to answer the alerts, in studies that could not rule out other causes; in randomized trials, alerts sent to a nurse manager, or even to a nurse with no other duties, did not cut deaths. A deterioration score belongs on a ward only where a staffed team receives its most urgent alerts directly and the hospital measures what that team achieves, and the medical staff, not the information technology department, should set its thresholds and check whom it misses.
Two randomized trials in St. Louis, published in 2013 and 2014 with a different, older model, show the limits of an alert. In the first, a text page to the ward's nurse manager whenever a patient crossed the alert threshold made no difference to how many were moved to intensive care or died. In the second, the alerts went instead to the rapid response team's nurse, a post with no other duties, who was to respond to each one within 20 minutes. Transfers and deaths still did not fall, though patients left the hospital about a day sooner. That trial was small and was designed to detect fewer transfers to intensive care, not fewer deaths.
The programs tied to fewer deaths built more around the alert, though the studies of them are observational. Kaiser Permanente's own program scanned the records of its Northern California hospitals every hour and sent each alert to a centralized team of specially trained nurses. They reviewed the chart off the ward, sparing bedside staff unnecessary interruptions, and called the unit's rapid response team when a patient needed to be seen; that team made a structured assessment and worked with the patient's physician. In a study by Kaiser's own researchers, about 16% of hospital stays with an alert were followed by a death within a month, against 20% of comparison stays that would have set one off had the system been running; adjusted for differences between the groups, the risk of death was about a sixth lower.
RWJBarnabas Health in New Jersey used Epic's index itself. At 11 of its hospitals, alerts at the highest-risk tier went automatically to rapid response teams unless rules suppressed them, and after the change fewer than half of high-risk stays produced a notification. In a before-and-after comparison by the system's own doctors and researchers at Rutgers, deaths among high-risk patients fell from about 23% to under 19%. The researchers credited a combination of factors, including staff education, the alerts and the notifications. Stephen O'Mahony, the system's chief medical information officer and the study's senior author, said the study had combined Rutgers' rigor with the system's reach: "The mortality benefit was not produced by an algorithm but by the partnership around the algorithm."
Stanford University Hospital used Epic's index differently, in an earlier version than the one now cleared. When a patient's score crossed 65, the alert went to the nurse and the physician at once, starting a structured huddle with a checklist. In a retrospective comparison of patients just above and just below that line, who differ little except in whether an alert fired, those above it were about 10 percentage points less likely to need a rapid response team, a move to intensive care or a code, though deaths did not fall.
Where the line sits decides how often a team is called, and Epic's own test on past records, submitted to the FDA, shows the trade. At the score's top threshold, about one flagged score in three was followed by a deterioration. Moving the line down to the lower threshold tripled the number of scores flagged, and the added scores, those between the two lines, were followed by a deterioration only about one time in eight. A lower line catches more of the patients who are about to decline, and calls the team far more often for patients who are not.
The line also decides whom the score misses. Pooled across Epic's test sites, its sensitivity at the lower threshold was about half for white patients, a third for Hispanic patients and a quarter for Black patients. In a site-by-site analysis the gap between Black and white patients largely closed. The bigger difference lay between the sites, and the one that supplied most of the data on Black patients was where the score caught the smallest share of deteriorations, in patients of every race.
The hardest objection comes from hospitals with the least to give: a staffed response asks the most of them, for a benefit the randomized trials did not find. A small hospital has no rapid response team to spare, and an imperfect warning on a nurse's screen seems better than none. The studies that found fewer deaths are observational, and answering costs staff: at RWJBarnabas, rapid response calls among high-risk patients rose by about half. The second St. Louis trial is the best evidence for this view, and it narrows the case: a staffed response has not been shown to be enough. But it gives no reason to prefer a bare alert, which changed nothing in the first trial, and it was not designed to detect a difference in deaths. Kaiser's program also put more around the alert: nurses off the ward screened every one before anyone on the ward was called, and a structured assessment on the ward brought in the patient's physician. Whether that made the difference is unproven, which is why each hospital should measure what its own response achieves. Nor must the answer sit on every ward: Kaiser's 21 hospitals shared a central team of nurses, and at Stanford the alert went to the patient's own nurse and physician, arrangements a small hospital could share or copy.
The principle that matters most is that every top-tier alert has an owner. Hospitals and health systems should switch the score on only where its most urgent alerts go straight to someone whose job is to answer them: a rapid response team, or a central team of nurses that screens alerts for several hospitals and calls in the ward's rapid response team, which brings in the patient's doctor. The extra calls need a budget before the first one comes, and a hospital that cannot staff or share such a response is better off leaving the alerts off.
Just as important is who draws the line. That choice belongs to the medical staff and nursing, through their patient safety committee, which can weigh what each setting costs in calls and missed patients and publish its choice. The same committee should track what happens to the patients the score flags, review every patient who went to intensive care, needed a code or died on a ward without an alert, and check whom the score misses by race, ethnicity and sex, as Epic's own submission did, and by unit. If the misses or the calls begin to harm patients, the threshold has to change, or the alerts have to go off. A physician asked to work alongside the score is entitled to three answers before it goes live: who receives its top alerts, who chose its thresholds and whom it misses.
The FDA's clearance settles what Epic's score is and what it may be used for. Whether it helps the next patient who begins to slip on a ward at night depends on who answers when it sounds, who decided where to set the line, and whether anyone is keeping count.
- What the FDA clearedFDA 510(k) K260008, Deterioration Index Version 2, applicant Epic Systems Corporation: received Jan. 2, 2026; decision Sept. 18, 2026, substantially equivalent; product code QNL, regulation 870.2210; traditional 510(k); predetermined change control plan authorized: no. The predicate is AgileMD's eCARTv5 Clinical Deterioration Suite (K233253). The 510(k) summary describes a static gradient boosted random forest (trained with LightGBM) that "uses 195 inputs" to predict "(1) escalation of care in the next 24 hours or (2) mortality in the next 72 hours," escalation meaning a transfer to the ICU, a rapid response team call or a code, in adults 18 or older on an inpatient unit, an observation unit or observation status in an emergency department. It is not for long-term care, operating rooms, post-anesthesia care units, labor and delivery or non-observation emergency patients. "Trained healthcare professionals at an organization determine the scores at which further patient evaluation should occur." Becker's Hospital Review reported the clearance on Oct. 8, 2026.
- Epic's validation, as submittedEpic's retrospective validation used data from "three geographically different sites in the United States," each one or more of academic hospitals, critical access facilities, research centers and safety-net organizations; no dates are given. In the main analysis, outcome prevalence was 3.8% (410,520/10,724,900) and the area under the ROC curve 0.83 (0.83-0.84), or 0.94 (0.94-0.94) for mortality alone. For all scores of 70 or more (the summary's "Medium (≥70)"): sensitivity 0.49, specificity 0.92, positive predictive value 0.20, flag rate 0.09. For scores of 90 or more: 0.28, 0.98, 0.34 and 0.03. Scores of 70 to 90 had a stratum-specific likelihood ratio of 3.78 (3.60-3.96); with the 3.8% prevalence, about 13% of them preceded a deterioration (a calculation from the summary's figures, not stated in it). Information from the device "is not intended to be relied upon as the sole basis for decisions that have an impact that may cause death or an irreversible deterioration of a person's state of health."
- Sensitivity by group and by siteIn the main analysis, for all scores of 70 or more, sensitivity was 0.51 (0.51-0.52) for White, 0.26 (0.24-0.30) for Black and 0.43 (0.40-0.47) for Other/Unknown patients; 0.32 (0.29-0.35) for Hispanic and 0.50 (0.49-0.51) for not Hispanic patients; 0.50 for female and 0.47 for male patients. Flag rates were 0.10 for White and 0.06 for Black patients. A second analysis (ICU stay or death within 24 hours, excluding observation patients) gave 0.48 (White), 0.31 (Black) and 0.35 (Hispanic). A "block aggregation analysis for the medium-risk threshold" by site, a term the summary does not define, gave sensitivity of 0.62 (White) and 0.63 (Black) at Site 1, 0.54 and 0.45 at Site 2, and 0.34 and 0.32 at Site 3; for Other/Unknown, 0.61, 0.58 and 0.44. Black patients accounted for 53.5% of blocks at Site 3, against 2.2% at Site 1 and 6.5% at Site 2. Becker's: "The gap between Black and white patients narrowed in a site-by-site analysis."
- RWJBarnabas Health: top alerts to rapid response teamsNahass and colleagues at RWJBarnabas Health and Rutgers (NEJM AI, published online July 29, 2026) evaluated the Epic Deterioration Index at 11 hospitals. Among 23,132 high-risk patients, deaths fell from 23.1% to 18.6%, "an 18% reduction in the risk-adjusted odds" of in-hospital death; rapid response activations rose from 25.3% of hospital stays to 37.5%. The researchers said the benefit "likely resulted from a combination of factors," including staff education and the automated notifications. According to 2 Minute Medicine's summary, scores of at least 60 triggered notifications "unless suppression criteria were applied" (for example, existing intensive or comfort care), and "Only 46.6% of post-intervention encounters generated notifications"; the adjusted odds ratio for death was 0.82 (95% CI, 0.74-0.91), against 0.86 (0.63 to 1.17) just above the threshold and 0.99 (0.70-1.39) just below it (scores of 55 to 59). Limitations include unequal seasonal distributions and incomplete implementation at two hospitals. Funding is not stated; the paper could not be opened.
- Kaiser Permanente: nurses who screen each alertKaiser Permanente's Advance Alert Monitor scanned the electronic health record hourly for significant risk of decline over the next 12 hours. Alerts went to a regional team of specially trained registered nurses, who reviewed the record off-site and, when on-site intervention was needed, contacted the unit's rapid response team, which performed a structured assessment and worked with the patient's physician. A New England Journal of Medicine study led by Gabriel Escobar of Kaiser's Division of Research covered the staggered deployment at 19 hospitals between Aug. 1, 2016, and Feb. 28, 2019 (all 21 had it by 2019), comparing 15,487 hospitalizations that reached the alert threshold with 28,462 comparison hospitalizations. Mortality within 30 days of an alert was 15.8% versus 20.4% unadjusted; the adjusted relative risk was 0.84 (95% CI, 0.78 to 0.90). Unadjusted ICU admission was 17.7% versus 20.9%, and length of stay among survivors 6.5 versus 7.2 days. Funded by the Gordon and Betty Moore Foundation and others.
- Stanford: an alert to the nurse and physician togetherGallo and colleagues (JAMA Internal Medicine, 2024;184(5):557-62) included 9,938 patients at Stanford University Hospital from January 2021 to November 2022. A score of 65 on the Epic Deterioration Index (range 0 to 100), described as an ordinal logistic regression model, triggered an alert to the patient's nurse and physician "to initiate collaborative measures, including a structured huddle and checklist." A retrospective regression discontinuity analysis of patients near the threshold ("a subset of 963 patients," per Stanford Medicine) found an absolute risk reduction of 10.4 percentage points in escalation of care (rapid response activation, ICU transfer or cardiopulmonary arrest; 95% CI, -20.1 to -0.8; P=0.03) and "no reduction in mortality," according to The Hospitalist's summary. Stanford Medicine (April 15, 2024) quotes the senior author, Ron Li: the intervention "is basically a conversation that otherwise may not have happened."
- Randomized trials of alerts, 2013 and 2014Bailey and colleagues (Journal of Hospital Medicine, 2013;8(5):236-242) ran a randomized, controlled crossover study on 8 general wards at an academic center from July 2007 to December 2011; alerts "were sent by text page to the nurse manager on intervention wards," and ICU transfer and death did not differ between groups. Kollef and colleagues (Journal of Hospital Medicine, 2014;9(7):424-429) randomized 571 patients on eight medicine units at Barnes-Jewish Hospital between Jan. 15 and May 9, 2013, sending alerts to the rapid response team's nurse, since 2012 "a dedicated position without other clinical responsibilities" instructed to respond within 20 minutes, or hiding them: ICU transfer 17.8% vs 18.2% (odds ratio 0.972; 95% CI 0.635-1.490), hospital mortality 7.3% vs 7.7% (0.947; 0.509-1.764), hospital duration 8.4±9.5 vs 9.4±11.1 days (P = 0.038). It was sized (514 patients) to detect an 8.0% absolute fall in ICU transfer from 20.0%. The trials share several authors.
- An outside comparison of the earlier indexEdelson and colleagues (JAMA Network Open, Oct. 15, 2024), a retrospective cohort study, included 362,926 adult medical-surgical ward encounters at seven Yale New Haven Health hospitals from 2019 to 2023; 16,693 encounters (4.6%) had a deterioration event, a transfer from a ward to an ICU or a death within 24 hours. The Epic Deterioration Index's area under the curve was 0.808, fifth of six (eCART 0.895; the free NEWS 0.829). At high-risk thresholds the researchers set at the specificity closest to a NEWS score of 7, median lead time was 11 hours for eCART, 8 for NEWS and 1 (IQR 0-39) for the Epic index; "The moderate-risk thresholds provided a median of at least 20 hours of lead time" for all six scores. Dr. Edelson reported equity interest in AgileMD and a patent for eCART; Dr. Churpek, a patent for eCART with royalties paid by the University of Chicago; no others were reported. Funded by the National Institutes of Health.