Half of all deaths in Romania's acute hospitals are recorded in 27 hospitals

What the tables show#

Four times a year, Romania's National Institute for Health Services Management (INMSS) publishes a set of tables on every hospital that reports stays to the National Health Insurance House. One of them says how stays ended: in death, in transfer to another hospital, or in discharge at the patient's request. The tables go down to ward level.

In 2025 hospitals reported 3,445,502 stays. 2.40 of 100 ended in death, about 82,692 deaths. Another 1.18 of 100 ended in transfer to another hospital and 3.0 of 100 in discharge at the patient's request.

The deaths are concentrated in few places. Half were recorded in 27 hospitals of the 431 for which the tables give the number of stays; those hospitals had 29% of stays. Almost all are county emergency hospitals and large emergency hospitals in Bucharest. At the other end, 82 hospitals with at least 2,000 stays each had 503,422 stays and about 422 deaths between them. In 18 of them, with 88,788 stays, the tables show no deaths at all.

A stay is an episode in hospital, not a person: a patient admitted three times is counted three times. The number of deaths is computed from the published percentages, rounded to two decimals, so it is approximate.

One specialty, two kinds of ward#

The differences stay very large between wards with the same name. In the 258 internal-medicine wards with at least 500 stays, the share of deaths runs from zero to 34 in 100. 22 wards, with 23,994 stays, reported no death in the whole year.

The difference goes together with the type of admission. In internal-medicine wards where under 30% of stays were emergency admissions, the median is 0.2 deaths per 100 stays. In wards where over 80% were, the median is 7.0.

The chart at the top shows the same tendency in most specialties: wards with many emergency admissions sit more often to the right, and wards with few sit more often next to zero. There are exceptions: a paediatric ward with nearly all stays admitted as emergencies can have no deaths, and an oncology ward with few emergencies can have many.

How much of the difference can be predicted#

I checked how much of these differences can be predicted without knowing anything about how patients are cared for. I fitted a model on the wards of 2023 and 2024 and had it predict the deaths of 2025, which it had not seen, in 2,956 wards.

The ward's specialty alone cuts the prediction error by 53% against a model that gives every ward the same share. Adding the share of emergency admissions brings the reduction to 70%. With the share of patients over 65, 73%. With the case-mix index and the share of stays admitted by transfer, 77%. This is an exploratory analysis, added after the hypotheses written in advance.

About a quarter stays unpredicted. It can hold the real severity of the cases, the way each hospital codes its patients, equipment, staffing and the quality of care. The published tables do not allow these to be separated.

The order of hospitals is stable: those that had more deaths than their ward mix predicts in 2023 are almost the same ones in 2025 (rank correlation 0.96). The tables do not show the cause of that stability.

A suspicion, tested, that did not hold#

Before comparing hospitals I wrote down a hypothesis and the threshold for calling it confirmed: hospitals with fewer deaths than their ward mix predicts transfer more patients to other hospitals or discharge more of them in a worsened state. Such a result would have been compatible with some deaths moving from one hospital to another.

The data point the other way. Among the 304 hospitals with at least 2,000 stays, the rank correlation between deaths and transfers plus worsened discharges is 0.23 (95% interval 0.11–0.34); the threshold written in advance was −0.20 or lower. Hospitals with more deaths than their ward mix predicts also have more transfers (0.29) and more discharges at the patient's request (0.31). For discharges in a worsened state the correlation is close to zero (0.06). For changes from 2019 to 2025 inside the same hospital the correlation is −0.10, with an interval that includes zero.

The test has a limit: more severe cases can raise deaths, transfers and discharges on request together, and that effect can mask any movement of deaths. The correct result is that no such movement can be seen in these tables. The tables cannot prove it does not happen.

A field that hospitals report very differently#

The same tables carry a "state at discharge": healed, improved, unchanged, worsened. Nationally, 19 of 100 stays are recorded as "healed" and 70 as "improved".

The order that sets out the patient record says only that the attending physician fills in the state at discharge. I found no criterion in its instructions separating "healed" from "improved" or "unchanged" from "worsened".

The reported shares differ widely. Of the 304 hospitals with at least 2,000 stays, 31 record "healed" for fewer than 1% of stays, and 12 for more than half. One hospital in ten records "healed" for at most 0.8% of stays, another one in ten for at least 37.9%. The tables cannot separate the mix of cases from the way each hospital fills in the field.

That is why I compare only three ways a stay can end that leave little room for interpretation: death, transfer to another hospital, and discharge at the patient's request.

The tables see one death in three#

In 2025 Romania registered 238,316 deaths (provisional). The deaths in reported hospital stays, about 82,692, are 35% of them. In 2019 the figure was 30%.

A study of 30 European countries for 2005–2017 found that the share of deaths taking place in hospital runs from 26% to 68%, with a median of 52%. The number computed here for Romania sits at the low end of that range, but it is not measured the same way: the study uses place-of-death statistics, and this page has only reported stays.

The rest do not appear in these tables: deaths at home, in care homes, in the emergency room before admission, and in units that do not report stays to the Insurance House. Deaths in the days after discharge appear only if the patient was admitted again and died in hospital.

In 2021, the second year of the pandemic, the share of stays ending in death reached 4.6 in 100, against 2.1 in 2019, and came down to 2.40 in 2025.

What should be published#

The numbers exist and are public, but a reader cannot learn from them whether a ward with 20 deaths per 100 stays cares for its patients worse than one with none. The difference can just as well come from the cases each one receives.

Other countries publish an indicator that takes the patients into account. In England, the health service publishes every month, for each hospital trust, the ratio between observed deaths and those expected from the characteristics of the patients, and it also counts deaths within 30 days of discharge. Its official definition says the indicator "is not a measure of quality of care"; it shows where a closer look is warranted. The Netherlands publishes a similar indicator, computed by the national statistics office.

In Romania, in-hospital mortality is one of the indicators in a hospital manager's contract. Until May 2023 the order asked for the "in-hospital mortality rate, for the whole hospital and for each ward". Since then it asks for the rate "for the whole hospital (caused by complications arising during the stay)". Neither wording takes account of the patients a hospital receives.

Romania collects the starting data: every stay is reported with diagnoses, procedures, age and type of admission, and deaths are recorded in the civil register. An adjusted indicator that includes deaths within 30 days of discharge would require linking the two sources; I have not checked how completely that can be done.

The 27 hospitals, of 431, where half of the deaths were recorded#