kondratevakate/sepsis-cases-event-log

Sepsis Cases Event Log

Anonymized hospital event log of ~1,050 patient trajectories suspected of sepsis, recorded from a hospital's ERP system. Each trajectory tracks admission, triage, lab orders (e.g. lactate, leucocytes), antibiotic administration, IV fluid administration, return-to-ER events, and discharge.

Widely used as a benchmark for process mining tasks: process discovery, conformance checking, predictive process monitoring.

Re-host notes

This Hugging Face repo is a lossless mirror of the original 4TU release. It ships the data in two forms:

• events.parquet — one row per event, flattened with PM4Py (pm4py.convert_to_dataframe). Drives the HF dataset viewer and is the fastest way to load the log for tabular ML.
• sepsis.xes.gz — byte-identical copy of the original 4TU file, for use with native process-mining tooling (PM4Py, ProM, Disco).

Shape

• 1,050 cases (patient trajectories)
• 15,214 events
• 16 activity types
• 39 attributes (case + event)

Usage

# Tabular access
from datasets import load_dataset
ds = load_dataset("kondratevakate/sepsis-cases-event-log", split="train")
print(ds.num_rows, ds.num_columns)

# Native XES via PM4Py
from huggingface_hub import hf_hub_download
import pm4py
path = hf_hub_download(
    "kondratevakate/sepsis-cases-event-log",
    "sepsis.xes.gz",
    repo_type="dataset",
)
log = pm4py.read_xes(path)

Schema notes

PM4Py convention is used in events.parquet:

• Case attributes are prefixed case: (e.g. case:concept:name is the case ID, plus case:Age, case:Diagnose, SIRS / infection flags, etc.).
• Event attributes use the original XES names: concept:name (activity), time:timestamp, org:group, lab measurements (Leucocytes, LacticAcid, CRP), etc.

Citation

@misc{mannhardt2016sepsis,
  author       = {Mannhardt, Felix},
  title        = {Sepsis Cases - Event Log},
  year         = {2016},
  publisher    = {4TU.ResearchData},
  version      = {1},
  doi          = {10.4121/uuid:915d2bfb-7e84-49ad-a286-dc35f063a460},
  url          = {https://doi.org/10.4121/uuid:915d2bfb-7e84-49ad-a286-dc35f063a460}
}

License

The upstream license is the 4TU General Terms of Use (https://data.4tu.nl/info/about-our-services/terms-of-use). The Hugging Face license tag is set to other so the upstream terms are represented faithfully; CC-BY-4.0 was considered but is not the actual upstream license. Attribution to Mannhardt (2016) is required.

Ethical considerations

Events and attribute values were anonymized in the original 4TU release. Downstream users remain responsible for compliance with applicable health-data regulations in their jurisdiction (e.g., HIPAA in the US, GDPR Article 9 in the EU, DHA HIPS in the UAE) and for obtaining any required institutional approvals (IRB) before research use.