Hugging Face's logo

Datasets:
SidneyBissoli
/
sih

Tasks:
Tabular Classification
Languages:
Portuguese
Size:
100M<n<1B
Tags:
health
brazil
public-health
parquet
datasus
sih
License:
Dataset card Files
xet
Community

The dataset is currently empty. Upload or create new data files. Then, you will be able to explore them in the Dataset Viewer.

SIH — Hospital Admission Records (Brazil, 1992–2026)

Individual-level hospital admission records from Brazil's Hospital Information System (SIH/SUS), covering 35 years of public hospital data. Each record corresponds to one Reduced Hospital Admission Authorization (AIH Reduzida — RD) and includes patient demographics, diagnoses (ICD-9 for 1992–1997, ICD-10 from 1998 onward), procedures, length of stay, costs, and outcome. Converted from legacy .dbc files to Apache Parquet.

Part of the healthbr-data project — open redistribution of Brazilian public health data.

Summary

Item Detail
Official source DATASUS FTP / Ministry of Health
Temporal coverage 1992–2026
Geographic coverage All 27 Brazilian states (by municipality of hospitalization)
Granularity Individual: one row per hospital admission (AIH)
Volume 415M+ records (11,011 .dbc files processed)
Format Apache Parquet, partitioned by ano/mes/uf
Data types All fields stored as string (preserves original format)
Update frequency Monthly (source publishes ~2–3 months after competency month)
License CC-BY 4.0

Resumo em português

SIH — Registros de Internações Hospitalares (Brasil, 1992–2026)

Microdados individuais de internações hospitalares do Sistema de Informações Hospitalares do SUS (SIH/SUS), cobrindo 35 anos de dados hospitalares públicos. Cada registro corresponde a uma Autorização de Internação Hospitalar Reduzida (AIH-RD) e inclui dados demográficos do paciente, diagnósticos (CID-9 para 1992–1997, CID-10 a partir de 1998), procedimentos, tempo de permanência, custos e desfecho. Convertidos de arquivos .dbc legados para Apache Parquet.

Item Detalhe
Fonte oficial FTP DATASUS / Ministério da Saúde
Cobertura temporal 1992–2026
Cobertura geográfica Todos os 27 estados brasileiros (por município de internação)
Granularidade Individual: uma linha por internação hospitalar (AIH)
Volume 415M+ registros (11.011 arquivos .dbc processados)
Formato Apache Parquet, particionado por ano/mes/uf
Atualização Mensal (fonte publica ~2–3 meses após o mês de competência)

Para documentação completa em português, consulte o repositório do projeto.

Data access

Data is hosted on Cloudflare R2 and accessed via S3-compatible API. The credentials below are read-only and intended for public use.

R (Arrow) 

library(arrow)
library(dplyr)

Sys.setenv(
  AWS_ENDPOINT_URL      = "https://5c499208eebced4e34bd98ffa204f2fb.r2.cloudflarestorage.com",
  AWS_ACCESS_KEY_ID     = "28c72d4b3e1140fa468e367ae472b522",
  AWS_SECRET_ACCESS_KEY = "2937b2106736e2ba64e24e92f2be4e6c312bba3355586e41ce634b14c1482951",
  AWS_DEFAULT_REGION    = "auto"
)

# Open a single partition (year/month/state)
ds <- open_dataset(
  "s3://healthbr-data/sih/ano=2024/mes=01/uf=SP/",
  format = "parquet"
)

# Example: hospital admissions in São Paulo, Jan 2024, by diagnosis
ds |>
  collect() |>
  count(DIAG_PRINC, sort = TRUE) |>
  head(20)

Important: Point to specific partitions (ano=YYYY/mes=MM/uf=XX/), not to the dataset root. The root contains README.md and manifest.json, which Arrow cannot read as Parquet files. You can also open broader paths like ano=YYYY/ or ano=YYYY/mes=MM/ to load multiple partitions at once.

Python (PyArrow) 

import pyarrow.dataset as pds
import pyarrow.fs as fs

s3 = fs.S3FileSystem(
    endpoint_override="https://5c499208eebced4e34bd98ffa204f2fb.r2.cloudflarestorage.com",
    access_key="28c72d4b3e1140fa468e367ae472b522",
    secret_key="2937b2106736e2ba64e24e92f2be4e6c312bba3355586e41ce634b14c1482951",
    region="auto"
)

# Single partition (year/month/state)
dataset = pds.dataset(
    "healthbr-data/sih/ano=2024/mes=01/uf=SP/",
    filesystem=s3,
    format="parquet"
)

# Example: admissions in São Paulo, Jan 2024
df = dataset.to_table().to_pandas()
print(df.head())
print(f"Records: {len(df)}, Columns: {len(df.columns)}")

Note: These credentials are read-only and safe to use in scripts. The bucket does not allow anonymous S3 access — credentials are required. Point to specific partitions, not the dataset root (see note above).

File structure 

s3://healthbr-data/sih/
  README.md
  manifest.json
  ano=1992/
    mes=01/
      uf=AC/
        part-0.parquet
      uf=AL/
        part-0.parquet
      ...
    mes=02/
      ...
  ...
  ano=2026/
    mes=01/
      ...

Historical schemas

The AIH form underwent major revisions from 1992 to 2015. The dataset preserves 14 distinct schemas:

Period Columns Key characteristics
1992–1993 35 Start of computerization; ICD-9; 2-digit year
1994 39
1995–1997 41–42 ICD-9; dates as YYMMDD
1998 41 Transition: ICD-10, 8-digit dates, 4-digit year
1999–2001 52–60 +ICU fields, +management fields
2002–2005 68–69 +CNES (2004)
2006–2007 75
2008–2010 86 FTP era change; SIGTAP 10-digit procedures; +RACA_COR
2011–2012 93
2013 95
2014–2026 113 Stabilized — +DIAGSEC1–DIAGSEC9

The number of columns varies by year. Columns not present in a given era will be absent from that partition's Parquet file. Use open_dataset(unify_schemas = TRUE) in Arrow to query across eras (missing columns filled with null).

Schema (modern era, 2014–2026, 113 columns)

Key variables in the most recent schema:

Variable Description
UF_ZI State code (processing)
ANO_CMPT Competency year
MES_CMPT Competency month
N_AIH AIH number (admission ID)
CNES Health facility code (CNES)
MUNIC_MOV Municipality of hospitalization (IBGE code)
MUNIC_RES Patient's municipality of residence (IBGE code)
NASC Patient's date of birth
SEXO Sex (1=male, 3=female)
IDADE Age
RACA_COR Race/color
DT_INTER Admission date
DT_SAIDA Discharge date
DIAS_PERM Length of stay (days)
DIAG_PRINC Primary diagnosis (ICD-10)
DIAGSEC1–DIAGSEC9 Secondary diagnoses (ICD-10)
PROC_REA Procedure performed (SIGTAP code)
VAL_TOT Total amount (R$)
MORTE Death during admission (0=no, 1=yes)
CEP Patient's ZIP code
CAR_INT Admission type
COMPLEX Complexity level

For the complete variable list across all 14 schemas, see the exploration document.

Source and processing

Original source: 11,011 .dbc files from the DATASUS FTP server, covering two directories: 200801_/Dados/ (modern era, 5,856 files) and 199201_200712/Dados/ (legacy era, 5,155 files). Scope: RD (AIH Reduzida) only — SP, RJ, ER file types are future expansions.

Processing: .dbc → R (read.dbc::read.dbc()) → all fields cast to character → Parquet (arrow::write_parquet()) → upload to R2 (rclone). No value transformations are applied — field values are published exactly as provided by the Ministry of Health. Dates in the legacy era (YYMMDD) are preserved as-is; ICD-9 codes (1992–1997) are not converted to ICD-10.

Bootstrap: Processed in 2 sprints: Sprint 1 (2008–2026, 5,856 files, 217.8M records, ~18h) + Sprint 2 (1992–2007, 5,155 files, 197.6M records, ~12–15h). Total: 11,011 files, 415,372,502 records, 16.1 GiB on R2.

Known limitations

  1. Government data, not ours. Values are preserved exactly as in the original .dbc files, including any inconsistencies or missing data.
  2. Fourteen historical schemas. The number of columns varies from 35 (1992) to 113 (2014+). Queries spanning multiple eras must handle missing columns.
  3. All fields are strings. Numeric fields (costs, age, length of stay) and dates must be parsed by the user.
  4. RD only. This dataset contains only AIH Reduzida (processed admissions). SP (professional services), RJ (rejected), and ER (errors) are not included.
  5. ICD-9 in legacy era (1992–1997). Diagnosis codes use ICD-9 (6 characters) before 1998 and ICD-10 (3–4 characters) from 1998 onward. No conversion is applied.
  6. Legacy date format (1992–1997). Dates use YYMMDD (6 digits) in the legacy era and YYYYMMDD (8 digits) from 1998 onward. No conversion is applied.
  7. 19 historical gaps. 19 files from the legacy era were not found on the FTP server (mostly Roraima 1995–2000, plus AC 1994 and AP 2007). These are Ministry-side gaps, not pipeline failures.
  8. Monthly partitioning. Unlike SINASC (annual), SIH is partitioned by year/month/state, reflecting the monthly publication frequency.

Related datasets

Dataset Period Records Link
SINASC (live births) 1994–2022 85M+ sinasc
SI-PNI Microdados (vaccination) 2020–present 736M+ sipni-microdados
SI-PNI COVID (vaccination) 2021–present 608M+ sipni-covid
SI-PNI Agregados — Doses 1994–2019 84M+ sipni-agregados-doses
SI-PNI Agregados — Cobertura 1994–2019 2.8M+ sipni-agregados-cobertura
SI-PNI Dicionários Static 263 rows sipni-dicionarios

Citation 

@misc{healthbrdata,
  author = {Sidney da Silva Bissoli},
  title  = {healthbr-data: Redistribution of Brazilian Public Health Data},
  year   = {2026},
  url    = {https://huggingface.co/datasets/SidneyBissoli/sih},
  note   = {Original source: Ministry of Health / DATASUS}
}

Contact

Last updated: 2026-03-09
Downloads last month
6