Analysis

run on 18-11-2026

Brazil's public real-world data: what it covers, how it links, what it can and can't answer

Brazil runs one of the largest single-payer health systems in the world, the SUS, and publishes the administrative trail of almost everything it does: every public hospital admission, every outpatient procedure it pays for, every death certificate, every live birth, every case of a notifiable disease. Most of it is record-level, national and free. For anyone who works with real-world data (RWD) in pharma, that is an unusual resource: a population of 213 million, three decades deep, with no licence fee.

It is also unusual in what it lacks. There is no patient identifier in any public file, so nothing links at person level. Coverage, coding and completeness change from source to source and from year to year. This piece maps what is there, measures how the pieces fit, and is explicit about what the data can and cannot support. Everything below was measured on a local copy of the official files in October 2026.

Heat strip: rows per year in each Brazilian public health data source, 1994–2026

The inventory

Twenty-four tables, 13 billion rows. Grouped by what they would be called in a pharma RWD team:

Source Years Records Unit of record Pharma analogue
SIH, hospital admissions 2009–2025 200 M one admission paid by SUS: ICD-10, dates, procedures, ICU days, amount paid, death in hospital inpatient claims, single public payer
SIH, professional services 2008–2025 2.4 B one billed professional act within an admission claims line items
SIA, outpatient production 2008–2025 6.2 B one outpatient procedure line; individual records (APAC) carry ICD-10 and cover oncology, dialysis, high-cost drugs outpatient and specialty-pharmacy claims
SIA, psychosocial care 2012–2025 138 M one month of community mental-health care outpatient episode records
SIM, deaths 1996–2022 31 M one death certificate: underlying and contributing causes (ICD-10), place, age, sex, race national death index
SINASC, live births 1994–2023 86 M one live-birth declaration: gestational age, birth weight, mother's age and education, congenital anomalies birth registry
SINAN, notifiable diseases 2000–2026 46 M one case report form: dengue, chikungunya, zika, yellow fever, malaria, severe acute respiratory infection, violence disease surveillance registries
SISVAN, nutrition 2008–2023 406 M one anthropometric visit in primary care EHR extract (vitals only)
SI-PNI, vaccination 1994–2019; 2020 microdata 94 M + 102 M doses by vaccine and municipality; one dose per record in the 2020 microdata immunisation registry
CNES, providers 2005–2025 950 M facility, bed and professional, monthly provider master file
ANS, private insurance 2014–2025 2.3 B beneficiary counts by plan type, municipality and month enrolment / eligibility file, aggregated
SNGPC, controlled drugs 2014–2021 10 M one pharmacy sale of a controlled drug, with prescriber state and ICD-10 pharmacy dispensing (partial copy)
Population 2000–2025 4.9 M residents by municipality, sex and age denominators

A few observations come straight out of the timeline. The registries are the oldest, from the mid-1990s. Billing data start in 2008–09. Each source closes at a different year: deaths at 2022, births at 2023, hospital billing at mid-2025, surveillance into 2026. A study period that needs all of them ends in 2022. The billing tables are the volume: 6.2 billion outpatient lines and 2.4 billion professional acts.

The copy of the controlled-drug sales file stops at exactly 10 million rows, with 2021 truncated. It is useful for exploring, not for counting.

Matrix: share of 2022 records whose municipality, facility and ICD-10 codes match the reference lists, by source

How it links: place, facility, diagnosis, time

No public file carries a patient identifier. SUS has one, the national health card number, but it is stripped before publication. Every cross-source analysis therefore links on four shared keys:

So linkage is ecological (municipality × period × diagnosis) or facility-level (what happens in a given hospital), never person-level. That rules out cohorts that follow a patient from diagnosis to death across systems. It still allows most descriptive and ecological designs, and facility-level studies of quality and volume.

The second lesson is less obvious: the same concept is coded differently in every system, and nothing fails when it is mixed up. Measured on 2022 records:

None of these raises an error. Each produces a result that looks plausible.

Heatmap: share of 2022 records with a usable value in each key field, by source

Fit for purpose, measured

In FDA and EMA terms, the question is not whether the data are good but whether they are good enough for a given question. Five dimensions, measured.

Completeness. Place, date, age and sex are close to complete in every source, with the exception noted above. Race/colour is the weak field: 75% in outpatient records, 80–85% in admissions, dengue notifications and nutrition records, 96–98% in deaths and births. Only a third of individual outpatient records carry an ICD-10 code. The bulk of outpatient production is reported in consolidated form, with no patient at all.

Heatmap: share of deaths with an ill-defined cause by state, 1996–2022

Cause-of-death quality. In 1996, 15.1% of deaths had an ill-defined underlying cause (ICD-10 R00–R99). By 2022 it was 6.1%, after a sharp fall in 2004–07 as northern and northeastern states set up death-investigation services. The fall was uneven: Rio de Janeiro went from 9.6% to 9.4% and is now third-worst. A cause-specific mortality comparison across states or decades that ignores this is partly measuring coding.

Bar chart: share of residents with a private medical plan by state, 2024

Coverage. The registries (deaths, births, notifications) cover everyone. The billing data cover only care paid by SUS. In December 2024, 52.2 million people, 24.6% of the population, held a private medical plan, from 4% in Roraima to 40% in São Paulo. Their admissions are invisible to the hospital data. Any hospital-based rate is therefore biased by state, and in the same direction as income. Private plan enrolment is published only as counts, so it can correct denominators but not identify patients.

Timeliness and lag. Billing data are processed monthly and arrive within months. An admission is billed in the month of discharge or later, so the billing year is not the year of care. Death-certificate microdata are released as final about two years after the year of death, and the latest year in this copy is 2022. Surveillance records are notified a median of 3 days after symptom onset, but the case is closed weeks later: the median time from notification to closure for dengue fell from 31 days in 2014 to 14 in 2025.

Stability over time. Classifications change under the analyst's feet. Dengue moved to the WHO 2009 case definitions in 2014, with new codes. From 2022 the public dengue file no longer includes discarded notifications, so "all notifications" means something different before and after. Comorbidity fields on the dengue form were blank until 2015 and usable from 2017. One state, Espírito Santo, reported dengue in its own system in 2021–2024 and almost disappears from the national file, while its hospitals kept billing dengue admissions.

Privacy. Hospital and surveillance records include date of birth, sex, municipality and, in hospital data, postcode. Combined, they can single out individuals in small municipalities. Everything published here is aggregated, and cells with fewer than 10 events are suppressed.

Worked example 1: infant mortality from two registries

Infant and neonatal mortality in Brazil 1996–2022 and by state in 2022

The infant mortality rate needs two registries: deaths under one year (SIM) divided by live births (SINASC). They share no identifier, but they share the residence municipality and the year, and that is enough for the rate. The direct ratio fell from 26.4 per 1,000 in 1996 to 11.6 in 2022. Neonatal deaths, in the first 28 days, depend more on hospital care. They fell by about half and are now two thirds of the total. Sergipe (16.8) and Amapá (16.5) are almost twice the Federal District (8.7).

The check is part of the example. In 2022 the death file has 30,791 records with age under one, 29,764 after removing 1,006 impossible negative ages and 21 records with no municipality. A second count by dates of birth and death gives 30,668. Where registration is weaker the direct ratio undercounts, which is why official estimates for the North and Northeast are corrected upward.

Worked example 2: dengue, from notification to hospital to death

Probable dengue cases, SUS admissions and deaths by year, 2014–2024

Dengue crosses three systems: notified in SINAN, admitted in SIH, certified in SIM. Between 2014 and 2024 surveillance recorded 16.7 million probable cases and 12,788 deaths. SUS hospitals billed 567,723 admissions for R$ 231.7 million. 2024 was off the scale: 6.4 million probable cases, 3.8 times the previous record, with 158,910 admissions and 6,270 deaths.

The series also shows how the hospital picture changed. In 2014 there was one SUS admission for every 16 probable cases; in 2024, one for every 40. In-hospital mortality among those admitted rose from 0.45% to 1.58%. Fewer patients reach a public bed, and those who do are sicker. That fits outpatient hydration of milder cases. It also fits more admissions moving to private hospitals, which SIH does not see. The data alone can't separate the two.

Deaths from dengue as recorded by surveillance, death certificates and SUS hospitals, 2014–2024

Three systems, three death counts. For 2014–2022, surveillance closed 5,422 cases as deaths, death certificates named dengue 5,925 times, and SUS hospitals recorded 2,755 in-hospital deaths. Each counts a different event, and without a patient identifier they cannot be reconciled one by one. A study that needs "dengue deaths" has to choose a definition and say which.

Within SINAN, which is record-level, risk factors can be read directly. In 2024 case fatality rose from 0.21 per 1,000 among children to 15.3 per 1,000 over age 80, and diabetic patients aged 60–79 died 2.9 times as often as non-diabetic ones of the same age. That comes with a caveat: comorbidity is recorded more carefully in severe cases.

Dengue serotype mix among typed cases, and share of probable cases typed, 2014–2026

The notification form is also the only place where a molecular result reaches the national file: the virus serotype, from RT-PCR or viral isolation. It is filled for very few cases, 0.2–0.8% of probable cases a year until 2021 and about 3% since 2023. Even so, it is enough to see the virus change. DENV-3 was almost absent for a decade, with fewer than 110 typed cases a year until 2023. Then it rose to 1,766 in 2024 and 6,857 in 2025, 14% of typed cases, and to 30% in the partial 2026 file. Meanwhile DENV-2 overtook DENV-1, from 30% of the mix in 2024 to 74% in 2025. A population that met only DENV-1 and DENV-2 has no immunity to DENV-3. That matters for an epidemic forecast, or for choosing trial sites for a vaccine or an antiviral. It is a signal, though, not a sample: which cases get typed depends on the state laboratory, not on the patient.

Worked example 3: high-cost drugs, by disease

SUS specialised-drug and cancer-drug authorisations by ICD-10, Brazil 2025

SUS dispenses high-cost medicines to outpatients through monthly authorisations (APAC). Each one names the procedure, which for drugs is the medicine, and the patient's diagnosis. In 2025 there were 50.5 million records from the Specialised Component of Pharmaceutical Care, the national programme for chronic and rare diseases. There were another 5.4 million for cancer drug therapy. Each record is one procedure for one patient in one month, so dividing by twelve gives a rough count of patients on treatment in an average month. Both files carry a diagnosis in every record, unlike the rest of outpatient billing. That makes this the closest public equivalent of a specialty-pharmacy claims file.

The diagnoses look like the treatment protocols. Lipid disorders, glaucoma, schizophrenia, chronic kidney disease, asthma and COPD lead by volume. Type 2 diabetes, rheumatoid arthritis, transplants and epilepsy follow. In cancer, breast tumours are 48% of all authorisations, most of them adjuvant therapy authorised month after month after surgery. Prostate cancer comes next, then myeloid leukaemia.

The amounts are the trap. For 7 of the 15 leading drug diagnoses, the bill for the whole year is R$ 0. The Ministry buys those medicines centrally and ships them to the states, so the authorisation records the dispensing but not the cost. In 2025 the authorisations billed R$ 2.03 billion. The federal budget line that buys these drugs spent R$ 12.85 billion in the same year. Between 2014 and 2020 the number of records rose by half while the billed amount fell by half. The authorisations measure how many patients get which treatment, where. They do not measure spending.

Growth of specialised-drug dispensing by diagnosis 2014–2025, and rheumatoid arthritis by state

Over time they show uptake. Between 2014 and 2025, records for psoriasis grew 6.9 times, lupus 5.4 times and ankylosing spondylitis 4.0 times, as new medicines, many of them biologics, entered the public treatment protocols. Rheumatoid arthritis grew 2.5 times. Inflammatory bowel disease doubled and multiple sclerosis grew 1.9 times. Alzheimer's disease barely moved. Across states the same disease is treated at very different rates. Paraná dispenses rheumatoid arthritis drugs to 309 patients per 100,000 in an average month, the national figure is 125 and Tocantins is at 30. Rheumatoid arthritis prevalence does not vary 10-fold between Brazilian states, so most of that gap is access to diagnosis and to the programme. For a launch plan or a market-access case, that gap is the finding. The usual limits apply. Without a patient identifier, a patient on two drugs counts twice, switching between drugs is invisible, and persistence cannot be measured.

The gap also shows why lineage matters. A first run of this section used a secondary copy of the outpatient file. In it, Alagoas's specialised-drug records fell from 582,000 in 2023 to 221,000 in 2024, and the state appeared at 13 patients per 100,000, a 22-fold gap with Paraná. Rebuilt from the Ministry's original files, Alagoas has 632,000 records in 2025 and 42 patients per 100,000. The drop was missing data, not missing patients, and nothing in the copy flagged it.

What it can and can't answer

Good fit:

Poor fit, or only with care:

Method

Data run on 5 October 2026; the serotype paragraph and worked example 3 on 18 November 2026.