---
license: mit
---
# Data Card: MARS Sepsis Blood Genomic Endotype Whole-Blood Microarray Dataset (GSE65682)

## Summary

Expression + sample metadata + feature metadata for GSE65682, the
Molecular Diagnosis and Risk of Sepsis (MARS) study, profiling the
whole-blood leukocyte transcriptome of critically ill ICU patients to
distinguish infectious sepsis (community-acquired pneumonia [CAP],
hospital-acquired pneumonia [HAP]) from non-infectious causes of critical
illness (no-CAP controls), and to classify patients into molecular
"endotypes." Profiled on the Affymetrix HG-U219 Array (GPL13667). PAXgene
blood RNA was collected at ICU admission and, for a subset, later in ICU
stay.

## Source accession

| Accession | N (samples) | Platform | Retrieval source |
|---|---|---|---|
| GSE65682 | 802 | Affymetrix HG-U219 (GPL13667), 24,646 probes | NCBI GEO |

This series bundles several independent, only partially-overlapping
sub-studies rather than one uniform cohort — each of the variables below
is populated for a different subset of the 802 samples, not a shared
disease-state axis:

- **Endotype/mortality** (479 samples: 263 discovery + 216 validation) — `endotype_cohort`, `endotype_class` (Mars1–4), `mortality_28d`, `time_to_event_28d`. Confirmed: every endotype-classified sample has a `mortality_28d` value.
- **ICU-acquired infection** (388 samples) — `icu_acquired_infection`, with a smaller paired subset (`icu_acquired_infection_paired`, 69 samples).
- **Pneumonia diagnosis** (225 samples: cap/hap/no-cap) — `pneumonia_diagnosis`.
- **Thrombocytopenia** (99 samples, graded A–D) — `thrombocytopenia`.
- **Diabetes mellitus** (410 samples) — `diabetes_mellitus`.
- **Abdominal sepsis / GI controls** (93 samples) — `abdominal_sepsis_or_control`.
- **Healthy controls** are flagged via `healthy_control` (logical).

Most fields are the literal string `"NA"` (or genuinely absent) outside
their relevant sub-study's sample set, not missing-at-random — treat each
variable's non-NA subset as its own analysis population.

## Files

- `sample_metadata.parquet` — one row per sample: `sample_id`, `title_identifier1`, `title_identifier2`, `healthy_control`, `sex`, `age`, `pneumonia_diagnosis`, `thrombocytopenia`, `endotype_cohort`, `endotype_class`, `mortality_28d`, `time_to_event_28d`, `icu_acquired_infection`, `icu_acquired_infection_paired`, `diabetes_mellitus`, `abdominal_sepsis_or_control`
- `feature_metadata.parquet` — one row per probe (`PROBEID`): annotation from `hgu219.db` (`SYMBOL`, `ENTREZID`, `ENSEMBL`, `GENENAME`, `UNIPROT`). Confirmed 100% probe-ID match against the package's keyspace (24,646/24,646).
- `expression.parquet` — single file, long format: `PROBEID`, `sample_id`, `value`

## Expression value processing

Values are **RMA-normalized, log2-transformed signal intensity**, as
deposited by the submitter. Confirmed already log-scale via the standard
detection heuristic; no further transformation applied.

## Provenance / reproducibility

See [pull_gse65682.R](pull_gse65682.R)