SidneyBissoli/sih

---
language:
- pt
license: cc-by-4.0
tags:
- health
- brazil
- public-health
- parquet
- datasus
- sih
- hospitalizations
- hospital-admissions
- morbidity
- sus
pretty_name: "SIH — Hospital Admission Records (Brazil, 1992–2026)"
size_categories:
- 100M<n<1B
task_categories:
- tabular-classification
source_datasets:
- original
---

# 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](https://huggingface.co/SidneyBissoli) 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](https://github.com/SidneyBissoli/healthbr-data).

## 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)

```r
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)

```python
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](https://github.com/SidneyBissoli/healthbr-data/blob/main/docs/sih/exploration-pt.md).

## 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](https://huggingface.co/datasets/SidneyBissoli/sinasc) |
| SI-PNI Microdados (vaccination) | 2020–present | 736M+ | [sipni-microdados](https://huggingface.co/datasets/SidneyBissoli/sipni-microdados) |
| SI-PNI COVID (vaccination) | 2021–present | 608M+ | [sipni-covid](https://huggingface.co/datasets/SidneyBissoli/sipni-covid) |
| SI-PNI Agregados — Doses | 1994–2019 | 84M+ | [sipni-agregados-doses](https://huggingface.co/datasets/SidneyBissoli/sipni-agregados-doses) |
| SI-PNI Agregados — Cobertura | 1994–2019 | 2.8M+ | [sipni-agregados-cobertura](https://huggingface.co/datasets/SidneyBissoli/sipni-agregados-cobertura) |
| SI-PNI Dicionários | Static | 263 rows | [sipni-dicionarios](https://huggingface.co/datasets/SidneyBissoli/sipni-dicionarios) |

## Citation

```bibtex
@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

- **GitHub:** [https://github.com/SidneyBissoli](https://github.com/SidneyBissoli)
- **Hugging Face:** [https://huggingface.co/SidneyBissoli](https://huggingface.co/SidneyBissoli)
- **E-mail:** [e-mail]

---

*Last updated: 2026-03-09*

### 数据集元数据
语言：
- 葡萄牙语（pt）
许可证：CC-BY 4.0
标签：
- 医疗健康
- 巴西
- 公共卫生
- Parquet
- DATASUS
- SIH
- 住院治疗
- 入院
- 发病率
- SUS
美观名称："SIH — 巴西住院入院记录（1992–2026）"
规模类别：
- 100M<n<1B
任务类别：
- 表格分类
源数据集：
- 原始数据集

# SIH — 巴西住院入院记录（1992–2026）

来自巴西医院信息系统（Hospital Information System, SIH/SUS）的个体级住院入院记录，涵盖35年的公立医疗机构数据。每条记录对应一份简化住院入院授权书（Reduzida Hospital Admission Authorization, AIH Reduzida — RD），包含患者人口统计学信息、诊断信息（1992–1997年使用ICD-9编码，1998年起使用ICD-10编码）、手术操作、住院时长、费用及转归。数据已从遗留的.dbc文件转换为Apache Parquet格式。

**隶属于[healthbr-data](https://huggingface.co/SidneyBissoli)项目**——巴西公共卫生数据开源再分发项目。

## 摘要
| 项目 | 详情 |
|------|--------|
| **官方数据源** | DATASUS文件传输协议（FTP）服务器 / 巴西卫生部 |
| **时间覆盖范围** | 1992–2026 |
| **地理覆盖范围** | 巴西全部27个州（按住院所在市政划分） |
| **数据粒度** | 个体级：每条记录对应一次住院入院（AIH） |
| **数据体量** | 超4.15亿条记录（已处理11011个.dbc文件） |
| **数据格式** | Apache Parquet，按`ano/mes/uf`（年/月/州）分区存储 |
| **数据类型** | 所有字段均存储为字符串（保留原始格式） |
| **更新频率** | 月度更新（数据源在统计月份结束后约2-3个月发布数据） |
| **许可证** | CC-BY 4.0 |

> 如需完整的葡萄牙语文档，请参阅[项目仓库](https://github.com/SidneyBissoli/healthbr-data)。

## 数据获取
数据托管于Cloudflare R2存储服务，通过兼容S3的API进行访问。以下凭据**仅具备只读权限**，面向公众开放使用。

### R语言（Arrow框架）
r
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"
)

# 打开单个分区（年/月/州）
ds <- open_dataset(
"s3://healthbr-data/sih/ano=2024/mes=01/uf=SP/",
format = "parquet"
)

# 示例：统计2024年1月圣保罗州的住院入院记录，按主要诊断分组
ds |>
collect() |>
count(DIAG_PRINC, sort = TRUE) |>
head(20)

> **重要提示：** 请指向具体的分区路径（`ano=YYYY/mes=MM/uf=XX/`），而非数据集根目录。根目录仅包含`README.md`和`manifest.json`，Arrow无法将其识别为Parquet文件。你也可以使用更宽泛的路径，如`ano=YYYY/`或`ano=YYYY/mes=MM/`，一次性加载多个分区。

### Python（PyArrow）
python
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"
)

# 加载单个分区（年/月/州）
dataset = pds.dataset(
"healthbr-data/sih/ano=2024/mes=01/uf=SP/",
filesystem=s3,
format="parquet"
)

# 示例：加载圣保罗州2024年1月的住院入院记录
df = dataset.to_table().to_pandas()
print(df.head())
print(f"记录数: {len(df)}, 字段数: {len(df.columns)}")

> **注意：** 上述凭据**仅具备只读权限**，可安全用于脚本中。该存储桶不支持匿名S3访问，必须使用凭据。请指向具体的分区路径，而非数据集根目录（详见上述提示）。

## 文件结构

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/
...

## 历史数据Schema
AIH表单在1992年至2015年间经历了多次重大修订。本数据集共保留了14种不同的Schema：
| 时间段 | 字段数 | 关键特征 |
|--------|:-------:|---------------------|
| 1992–1993 | 35 | 计算机化起步阶段；使用ICD-9编码；年份为2位数字 |
| 1994 | 39 | |
| 1995–1997 | 41–42 | 使用ICD-9编码；日期格式为YYMMDD |
| 1998 | 41 | **转型阶段：** 采用ICD-10编码，日期为8位数字，年份为4位数字 |
| 1999–2001 | 52–60 | 新增ICU相关字段、管理类字段 |
| 2002–2005 | 68–69 | 新增CNES编码（2004年起） |
| 2006–2007 | 75 | |
| 2008–2010 | 86 | **FTP服务器格式变更；** 采用SIGTAP 10位操作编码；新增`RACA_COR`（种族/肤色）字段 |
| 2011–2012 | 93 | |
| 2013 | 95 | |
| 2014–2026 | 113 | **格式稳定阶段：** 新增`DIAGSEC1`–`DIAGSEC9`（次要诊断）字段 |

不同年份的字段数量存在差异。特定时期不存在的字段将不会出现在对应分区的Parquet文件中。在Arrow中使用`open_dataset(unify_schemas = TRUE)`可实现跨时期查询（缺失字段将以`null`填充）。

## 最新Schema（2014–2026年，共113个字段）
最新Schema中的关键变量如下：
| 字段名 | 说明 |
|----------|-------------|
| `UF_ZI` | 州代码（处理用） |
| `ANO_CMPT` | 统计年度 |
| `MES_CMPT` | 统计月份 |
| `N_AIH` | AIH编号（入院ID） |
| `CNES` | 医疗机构代码（CNES） |
| `MUNIC_MOV` | 住院所在市政（IBGE代码） |
| `MUNIC_RES` | 患者常住市政（IBGE代码） |
| `NASC` | 患者出生日期 |
| `SEXO` | 性别（1=男性，3=女性） |
| `IDADE` | 年龄 |
| `RACA_COR` | 种族/肤色 |
| `DT_INTER` | 入院日期 |
| `DT_SAIDA` | 出院日期 |
| `DIAS_PERM` | 住院时长（天） |
| `DIAG_PRINC` | 主要诊断（ICD-10编码） |
| `DIAGSEC1`–`DIAGSEC9` | 次要诊断（ICD-10编码） |
| `PROC_REA` | 执行的手术操作（SIGTAP编码） |
| `VAL_TOT` | 总费用（巴西雷亚尔，R$） |
| `MORTE` | 住院期间死亡（0=否，1=是） |
| `CEP` | 患者邮政编码 |
| `CAR_INT` | 入院类型 |
| `COMPLEX` | 医疗复杂度等级 |

> 如需获取全部14种Schema的完整字段列表，请参阅[探索文档](https://github.com/SidneyBissoli/healthbr-data/blob/main/docs/sih/exploration-pt.md)。

## 数据源与处理流程
**原始数据源：** 来自DATASUS FTP服务器的11011个.dbc文件，涵盖两个目录：`200801_/Dados/`（现代时期，共5856个文件）和`199201_200712/Dados/`（遗留时期，共5155个文件）。本数据集仅包含RD（AIH Reduzida）类型数据——SP、RJ、ER文件类型将在后续版本中补充。
**处理流程：** .dbc文件 → 使用R语言的`read.dbc::read.dbc()`读取 → 所有字段转换为字符型 → 写入Apache Parquet格式（`arrow::write_parquet()`） → 通过`rclone`上传至R2存储。未对字段值进行任何转换——所有字段值均严格保留巴西卫生部提供的原始格式。遗留时期（1992–1997年）的日期格式保留为YYMMDD；1992–1997年的ICD-9编码未转换为ICD-10。
**处理进度：** 分两阶段完成：第一阶段（2008–2026年，5856个文件，2.178亿条记录，耗时约18小时）+ 第二阶段（1992–2007年，5155个文件，1.976亿条记录，耗时约12–15小时）。总计处理11011个文件，共415,372,502条记录，在R2存储中占用16.1 GiB空间。

## 已知局限性
1. **政府原始数据，未经修改：** 所有字段值均严格保留原始.dbc文件中的内容，包括任何不一致或缺失数据。
2. **14种历史Schema：** 字段数量从1992年的35个到2014年及以后的113个不等。跨时期查询需处理缺失字段的情况。
3. **所有字段均为字符串类型：** 数值字段（费用、年龄、住院时长）和日期格式需由使用者自行解析。
4. **仅包含RD类型数据：** 本数据集仅包含AIH Reduzida（已处理的入院记录）。SP（专业服务）、RJ（被驳回记录）和ER（错误记录）类型数据未包含在内。
5. **遗留时期使用ICD-9编码（1992–1997年）：** 1998年之前的诊断编码使用ICD-9（6位字符），1998年起使用ICD-10（3–4位字符），未进行编码转换。
6. **遗留时期日期格式（1992–1997年）：** 遗留时期的日期格式为YYMMDD（6位数字），1998年起为YYYYMMDD（8位数字），未进行格式转换。
7. **19处历史数据缺失：** FTP服务器上缺失19个遗留时期的文件（主要为1995–2000年的罗赖马州数据，以及1994年的阿克里州和2007年的阿马帕州数据）。此类缺失为卫生部侧的数据缺口，非数据处理流程故障。
8. **按月分区：** 与SINASC数据集（按年分区）不同，本数据集按年/月/州分区，以匹配数据源的月度发布频率。

## 相关数据集
| 数据集 | 覆盖时段 | 记录数 | 链接 |
|---------|--------|---------|------|
| SINASC（活产记录） | 1994–2022 | 超8500万条 | [sinasc](https://huggingface.co/datasets/SidneyBissoli/sinasc) |
| SI-PNI Microdados（疫苗接种） | 2020年至今 | 超7.36亿条 | [sipni-microdados](https://huggingface.co/datasets/SidneyBissoli/sipni-microdados) |
| SI-PNI COVID（新冠疫苗接种） | 2021年至今 | 超6.08亿条 | [sipni-covid](https://huggingface.co/datasets/SidneyBissoli/sipni-covid) |
| SI-PNI 聚合数据 — 接种剂次 | 1994–2019 | 超8400万条 | [sipni-agregados-doses](https://huggingface.co/datasets/SidneyBissoli/sipni-agregados-doses) |
| SI-PNI 聚合数据 — 接种覆盖率 | 1994–2019 | 超280万条 | [sipni-agregados-cobertura](https://huggingface.co/datasets/SidneyBissoli/sipni-agregados-cobertura) |
| SI-PNI 词典文件 | 静态 | 263行 | [sipni-dicionarios](https://huggingface.co/datasets/SidneyBissoli/sipni-dicionarios) |

## 引用格式
bibtex
@misc{healthbrdata,
author = {Sidney da Silva Bissoli},
title = {healthbr-data: 巴西公共卫生数据再分发项目},
year = {2026},
url = {https://huggingface.co/datasets/SidneyBissoli/sih},
note = {原始数据源：巴西卫生部 / DATASUS}
}

## 联系方式
- **GitHub：** [https://github.com/SidneyBissoli](https://github.com/SidneyBissoli)
- **Hugging Face：** [https://huggingface.co/SidneyBissoli](https://huggingface.co/SidneyBissoli)
- **电子邮箱：** [e-mail]

**最后更新时间：** 2026年3月9日