Last-resort antibiotics, hospital by hospital

Reported: 1.5%. Estimated for Europe: 6.5%

The rate hospitals publish is simple: patients with a hospital-acquired infection, divided by patients discharged, times a hundred. The rates of the 100 hospitals that also published a usable number of discharges add up to roughly 17,136 patients with an infection out of 1,131,381 discharged: 1.5%. 68 of the 107 hospitals reported under 1%.

That matches Romania's national reporting: 1.26% in 2023, according to the National Institute of Public Health (INSP). The same report sets it beside the European estimate, 6.5% of discharged patients, from the European Centre for Disease Prevention and Control (ECDC) surveys of 2022–2023, and writes of under-reporting.

Percent of discharged patients. Sources: hospital annexes (2025), CARMIAAM 2023, p. 62.

One clue why: in the ECDC survey Romanian hospitals sent 14.6 blood cultures to the lab per thousand patient-days, against a European median of 44.7. An infection nobody tests for never reaches the statistics. On the survey day 42.9% of patients in Romanian hospitals were on antibiotics, and Romania's antibiotic-resistance index was the highest of the countries surveyed.

The test finds no sign that hospitals buying more reserve antibiotics report fewer infections

Meropenem, imipenem, ertapenem, colistin and tigecycline are the antibiotics held back for bacteria that have learnt to resist the others, many of them caught in hospital. A hospital that buys a lot of them and reports almost no infections deserves a question. The test was written and locked before any calculation, with a stopping rule: fewer than 60 hospitals with both figures, and the investigation stopped.

70 hospitals qualified. Direct purchases in 2025, per bed, say almost nothing about the reported rate: the rank correlation is 0.03 (95% confidence interval −0.20 to 0.24). The variant that adds framework contracts, as an estimate of annual capacity from awards published in 2024–2025, gives 0.16 (−0.09 to 0.39) on the same hospitals, and 0.21 (0.01 to 0.41) on the wider group of 88 hospitals where it can be worked out. No version has the sign the hypothesis predicts, although the intervals do not rule out a weak negative link.

No interval by type excludes zero: the groups are small, 13 to 22 hospitals.

A weak correlation can still hide a few cases. So the main test was a list: hospitals in the fifth that buys the most per bed and, at the same time, in the fifth that reports the least, or zero.

At the edge of the test: two town hospitals

The list has two names. Făget town hospital, in Timiș county, with 115 beds, bought meropenem, ertapenem, colistin and tigecycline by direct purchase in 2025, enough for 238 treatment-days, in 12 purchases. For the whole year it reported a rate of 0.14%: fewer than ten patients with a hospital-acquired infection, out of 3,764 discharged.

Corabia town hospital, in Olt county, with 123 beds, bought meropenem for 233 treatment-days, more than half of it in the last quarter (in the fourth quarter, 40 boxes of ten one-gram vials), and also reported 0.14% for 2025, out of 3,382 discharged. The quantities come from the procurement system's item lines; the test used the estimate from value, 227 days.

Both sit exactly on the test's thresholds. The plan asked for the bottom fifth without saying how to compute the threshold; the code that ran ends it at the 15th of the 70 rates: 0.14%. Făget and Corabia are 15th and 16th. With an interpolated threshold (0.138%) the list would be empty. Corabia also sits exactly on the purchase threshold, 1.84 treatment-days per bed. The framework variant names no hospital.

What these numbers do not prove: a treatment-day bought may never have been used; meropenem also treats infections caught at home and is sometimes given to prevent infection during surgery; a small hospital may take in transferred patients who already have an infection. The distance between what was bought and what was reported is a question for each hospital. The page implies no motive.

A whole year without a single infection

Four hospitals printed zero for all of 2025. Three of them still bought meropenem that year: Lipova town hospital and Sebiș chronic-care hospital, both in Arad county, and Rovinari town hospital, in Gorj. The quantities are small, 12 to 17 treatment-days, and a chronic-care hospital may treat infections caught at home. Lipova and Rovinari print zero on every ward. A whole year without a single infection is still a rare result.

The fourth, Cisnădie town hospital, in Sibiu county, printed zero in its official annex. Its own infection-prevention unit's note for 2025, registered on 9 January 2026 and published on the same website, counts ten infections, an incidence of 0.67%. The two documents cannot both be right. In October 2025 Cisnădie also bought four boxes of meropenem.

The documents: Cisnădie annex · infection-prevention note

A second source: Bucharest

Bucharest's Public Health Directorate published, for 2024, the number of infections reported by the city's 42 public healthcare units with beds. The year differs, so this is a year-to-year plausibility check, not a confirmation. For seven of eight comparable hospitals, the number worked out from the 2025 annex lies between half and double the 2024 count. The exception is the Alessandrescu-Rusescu mother-and-child institute: 213 infections in 2025, almost six times the 37 of 2024.

The planned same-year check, against county totals for 2025, was not possible: no such totals were found. Both sources may share the same omissions.

Limits

• Direct purchases are only part of what a hospital buys. Tenders and orders under framework contracts carry no quantities in the open data, so the direct-purchase total is a floor. County hospitals buy almost everything through frameworks.
• Framework contracts appear in EU notices with the winning bid's value, not the quantity ordered. The page turns them into estimated annual capacity: the largest winning bid of each lot, divided by the contract's length. Repeats of a lot across notices are removed by title and value, without reconciling procedures: an amended value may survive as a separate lot, and two procedures with the same title and value are counted once. Bids may be multi-year ceilings.
• Bought is not used. A hospital may stock the drugs or pass them to another hospital.
• Carbapenems and colistin also treat infections caught outside hospital. In Romania 63.5% of surgical prophylaxis courses last more than a day (ECDC), which adds use without any infection.
• The published rate counts patients, not infections, and depends on how hard a hospital looks. A low rate may mean few infections or few tests.
• Treatment-days are World Health Organization defined daily doses (meropenem 3 g, imipenem 2 g, ertapenem 1 g, colistin 9 million units, tigecycline 100 mg). A real patient may get a different dose.
• Where the infection count worked out from a rate comes to fewer than ten, the page says "fewer than ten". The hospital's printed rate is a public document and is shown as published.
• Beds are continuous-care beds from the annex. The form does not ask for patient-days, so the page cannot report per hundred patient-days as the plan intended.
• Cards and the table show item-line quantities where they were read, including for the named hospitals. The registered test uses the quantities fixed before the calculation; dots on the chart sit where the test put them.

Method

The list: the 191 public hospitals that appear either in 2025 direct purchases naming one of the five antibiotics, or in 2024–2025 TED award notices. For each, the 2025 annex was sought on its website, downloaded and read; rate, beds and discharges were copied exactly, with the page and the file's fingerprint. A second, blind reading of 20 randomly chosen annexes gave the same rate in all 20. Hospitals not found on the first pass were searched again by other readers; nine more rates turned up that way.

Purchases: 815 direct purchases in the 2025 SEAP files naming one of the five antibiotics, minus the lab discs and tests that carry the same names. Quantities come from SICAP.ai item lines where they could be read; otherwise from value divided by the median price of the same product. On 50 purchases checked both ways, the median ratio of estimate to item-line quantity was 1.00, and 39 of 50 came within a quarter.

The plan, with its tests and stopping rule, was fixed and fingerprinted before any calculation. Departures are set out in the project folder. Three code errors were fixed after the first run: a "1,128" read as 1 for beds; hospital type; and, in the framework variant, several suppliers' bids added up for the same lot (the fix removes a third hospital from that variant's list). The method review found two more: orders with several drugs counted only the first, and 192 of the EU lots were lab products (discs, tests, strips). Re-run with corrections, the main test gives 0.03 and the same list; among hospitals with no framework contract, −0.01 (registered: −0.01). The page shows the corrected framework variant; that variant's registered result, before lab products were removed, was 0.23 on 87 hospitals and 0.15 on the same 70.

The plan grouped institutes with county and teaching hospitals; the calculation put them with specialist hospitals. With the planned grouping, the correlations by type are 0.29 for county, teaching and institute hospitals (16) and −0.12 for specialist ones (19); no interval excludes zero. The plan's descriptive test, treatment-days bought directly per reported infection, has a median of 3.4 across 66 hospitals; it is not shown per hospital because many have fewer than ten infections. The plan said cells under ten get no rate; the page shows the hospital's printed rate and hides only worked-out numbers under ten. With the nine rates found in the second search, the test gives 75 hospitals and 0.03; the list stays the same. At Elena Doamna in Iași, the test's plausibility filter dropped the 16,928 discharges printed in its annex; the descriptive figures use them.

Download the data (CSV, with dictionary) Data dictionary

The data can be rebuilt with one command from the project folder. The next update makes sense once the 2026 annexes are out, at the end of January 2027.

Hospitals named on this page may send corrections or explanations; they are published here with the date received.

From the same series

• Protection orders, court by court
• Drink-driving: what each court hands down
• What Romania pays at the European Court of Human Rights

Sources

• Health Minister's Order 1117/2025 (Official Gazette 913, 3 October 2025), the annex every public hospital must publish
• The annexes on the websites of the 191 hospitals; each one is linked from its hospital card and in the download
• 2025 direct purchases: the quarterly SEAP files on data.gov.ro (OGL-ROU 1.0 licence)
• Item lines (quantity, price) for a sample of direct purchases: SICAP.ai
• 2024–2025 award notices for framework contracts: Tenders Electronic Daily (TED), European Union
• INSP–CNSCBT, Antibiotic consumption, antimicrobial resistance and healthcare-associated infections, Romania 2023 (CARMIAAM), pp. 6 and 62
• ECDC, Point prevalence survey of healthcare-associated infections and antimicrobial use in European acute care hospitals 2022–2023, Romania country sheet
• Bucharest Public Health Directorate, Activity report 2024, pp. 12–15
• WHO ATC/DDD Index 2025 (defined daily dose)