sih-br-mcp, microdatasus or PySUS? One question about Brazilian hospital admissions, three ways

Pneumonia among Espírito Santo residents in 2024: the microdata in R, the microdata in Python and an MCP server inside an AI assistant — the same total, what each one costs, and a trap in the middle

Public health

DataSUS

SIH

MCP

microdatasus

PySUS

Author

Sidney Bissoli

Published

November 2, 2026

Published 2 November 2026. Every number here was measured that day: sih-br-mcp 1.8.1 on the public endpoint, microdatasus 3.0.0 with dplyr 1.2.1 and R 4.6.1 on Windows, PySUS 2.11.7 with Python 3.13 on Linux, over a residential connection in Brazil. The scripts are in scripts/comparativo/ of the sih-br-mcp repository; nothing is illustrative. The same comparison, as a product page: sih.sidneybissoli.com/vs-microdatasus-pysus.

If you work with Brazilian public hospital admissions — SIH/SUS, the AIH records DATASUS publishes every month — you almost certainly reach the data through one of two packages: microdatasus, by Raphael Saldanha, in R, or PySUS, by the AlertaDengue team, in Python. Both download and decode the microdata: one record per admission. sih-br-mcp is mine and sits one layer up: an MCP server that answers the aggregate question inside an AI assistant (Claude, ChatGPT, any MCP client) from pre-built cubes, with provenance in every answer.

They are tools for different jobs, which is exactly why it is worth putting all three in front of the same question.

The question, and why it needs 432 files

How many pneumonia admissions (ICD-10 J12–J18) did residents of Espírito Santo have in 2024, and how many deaths?

It looks like one file. DATASUS publishes SIH as one file per hospital state and billing month (the month the AIH was billed, not the month of admission). The question is by state of residence and admission date. Two consequences:

A resident of Espírito Santo admitted in Minas Gerais is in Minas Gerais’s file.
An admission in December 2024 may be billed in March 2025; the four following billing months close 99.7–99.9% of a year.

The honest answer from microdata: 27 states × 16 billing months (2024-01 to 2025-04) = 432 files, 1,358 MiB of .dbc, 18,863,265 records.

The natural shortcut, reading only Espírito Santo’s 16 files (30 MiB), gives 13,097 admissions. It misses 208: residents admitted in Minas Gerais (102), Rio de Janeiro (83) and elsewhere. Small (1.6%), and exactly the kind of difference nobody notices.

Three ways

	Admissions	Deaths	Downloaded	Time
sih-br-mcp 1.8.1	13,305	1,250	21 KB response	0.9 s
microdatasus 3.0.0	13,305	1,250	432 files, 1,358 MiB	40.6 min
PySUS 2.11.7, `group="RD"`	9,135	760	345 of 432 files	59.4 min*

* On Linux under WSL, reading and writing the Windows disk: not comparable with microdatasus’s time.

microdatasus and sih-br-mcp reach the same number, group by group: J12 461, J13 37, J14 1, J15 3,648, J16 265, J17 100, J18 8,793. That is neither coincidence nor anyone’s merit: sih-br-mcp’s cubes are built from a 1:1 Parquet mirror of the same .dbc files, over the same billing-month window. What differs is where the work happens. With microdata, on your machine, for every question. With sih-br-mcp, once, when the cubes are built; the question only reads the result.

PySUS fell short for a reason that gets its own section below.

microdatasus (R)

One file at a time: binding all 432 does not fit in memory.

library(microdatasus)
library(dplyr)

ufs   <- c("AC", "AL", "AM", "AP", "BA", "CE", "DF", "ES", "GO", "MA", "MG", "MS", "MT", "PA",
           "PB", "PE", "PI", "PR", "RJ", "RN", "RO", "RR", "RS", "SC", "SE", "SP", "TO")
comps <- seq(as.Date("2024-01-01"), as.Date("2025-04-01"), by = "month")

pneu <- list()
for (d in as.list(comps)) for (uf in ufs) {
  y <- as.integer(format(d, "%Y")); m <- as.integer(format(d, "%m"))
  rd <- fetch_datasus(year_start = y, month_start = m, year_end = y, month_end = m,
                      uf = uf, information_system = "SIH-RD",
                      vars = c("MUNIC_RES", "DT_INTER", "DIAG_PRINC", "MORTE"))
  pneu[[length(pneu) + 1]] <- rd |>
    filter(substr(MUNIC_RES, 1, 2) == "32",                     # lives in ES
           substr(DT_INTER, 1, 4) == "2024",                    # admitted in 2024
           substr(DIAG_PRINC, 1, 3) %in% sprintf("J%02d", 12:18))
}
bind_rows(pneu) |> summarise(admissions = n(), deaths = sum(MORTE == "1"))
#>   admissions deaths
#> 1      13305   1250

432 files, no errors, 40.6 minutes. fetch_datasus()’s vars = helps: only four columns stay in memory. Note MORTE == "1": fetch_datasus() returns everything as character, and that is what you compare — after process_sih(), the code becomes "Sim"/"Não".

PySUS (Python)

import pandas as pd
import pysus

ufs = "AC AL AM AP BA CE DF ES GO MA MG MS MT PA PB PE PI PR RJ RN RO RR RS SC SE SP TO".split()
comps = [(2024, m) for m in range(1, 13)] + [(2025, m) for m in range(1, 5)]
pneumonia = {f"J{n}" for n in range(12, 19)}

parts = []
for y, m in comps:
    for uf in ufs:
        rd = pysus.ftp.sih(state=uf, year=y, month=m, group="RD", as_dataframe=True)
        if rd.empty:
            continue
        parts.append(rd[rd["MUNIC_RES"].str[:2].eq("32")
                        & rd["DT_INTER"].str[:4].eq("2024")
                        & rd["DIAG_PRINC"].str[:3].isin(pneumonia)])
pneu = pd.concat(parts)
len(pneu)
#> 9135

sih-br-mcp (inside the assistant)

The question goes in plain language. The assistant picks the tool, and the call it makes is this:

get_hospitalizations {
  "uf": ["ES"], "year": [2024], "cid_chapter": [10], "group_by": ["cid_group"]
}

The answer carries one row per ICD group (J12 to J18 add up to 13,305 admissions and 1,250 deaths) and a provenance block saying where each number came from: the billing months read (2024-01 to 2025-04, all 27 file states), the mirror’s vintage, the MD5 of every source .dbc and the formal DATASUS citation. First call in the session, 0.9 s; the second, 0.7 s.

The trap in the middle: PySUS 2.11.7 and the files without a group

With group="RD", the loop above returned 9,135 admissions, not 13,305 — 31% fewer, with no error and no warning.

PySUS 2.x does not list the DATASUS FTP on each call: it queries its own catalog (a DuckDB database downloaded on first use) pointing to Parquet copies of the files. In that catalog, 4,526 of the 10,946 SIH RD files have no group. A group="RD" query filters by group and does not see them. In this question’s window that is 87 of 432 files — including 7 of Espírito Santo’s own 16 months (Jan, Feb, Mar, Jul and Nov 2024; Jan and Apr 2025) — and the pneumonia admissions inside them add up to 4,170. The arithmetic closes: 9,135 + 4,170 = 13,305.

Calling without group reaches those files, together with the month’s other groups (ER, RJ and SP), which you then drop by name. On Windows there is a second defect, upstream of that one: the catalog returns WindowsPath objects, their str() comes out with backslashes, and a prefix check written with forward slashes discards them all. The SIH calls came back empty, and so did the file listings for SIM and SINASC. That one had already been reported and is fixed in the code since 30 October (PR #374), awaiting the next release.

The missing-group one I reported the same day. It is the kind of defect fixed in an afternoon. Meanwhile, anyone using PySUS 2.11 for SIH should check how many files came back, not only how many rows.

Volume, in numbers

	Size
Raw SIH-RD microdata, 1992–2026, in the mirror	23.86 GiB, 11,211 files, 432,173,746 records
Everything sih-br-mcp’s cube channel publishes	1.63 GiB
This post’s question	1,358 MiB from microdata; 21 KB from sih-br-mcp
Admissions, days, spending and deaths per year, 1992–2025	420,107,114 admissions, one call under 1 s

What sih-br-mcp does not do

Individual records. The cubes are aggregates; no AIH can be recovered.
Variables the cubes do not carry: procedure performed, facility (CNES), admission type, secondary diagnosis.
Other DATASUS systems: SIM, SINASC, SIA, SINAN, CNES.
What DATASUS published today. The vintage is the mirror’s, and every answer says which.

For all of that, use microdatasus or PySUS.

Which one

Aggregate question, auditable answer, inside the assistant: sih-br-mcp.
Record-level data, variables outside the cubes, other systems, in R: microdatasus (and read.dbc if you only need to open the .dbc).
The same, in Python: PySUS, checking how many files came back.
To check sih-br-mcp: redo the count from microdata, as this post did. A difference is a bug; open an issue with the slice.

The three complement more than they compete. sih-br-mcp exists only because the microdata is open and because microdatasus and PySUS showed, for years, what can be done with it. Thanks to Raphael Saldanha and the PySUS team.