---
license: cc-by-4.0
pretty_name: Synthetic Clinical Telemetry for Early-Warning Scoring
language:
  - en
tags:
  - synthetic-data
  - healthcare
  - clinical-monitoring
  - time-series
  - tabular
  - early-warning-scores
  - news2
task_categories:
  - tabular-classification
size_categories:
  - 1K<n<10K
configs:
- config_name: patients
  data_files:
  - split: train
    path: patients.parquet
- config_name: telemetry_vitals
  data_files:
  - split: train
    path: telemetry_vitals.parquet
- config_name: telemetry_edge_cases
  data_files:
  - split: train
    path: telemetry_edge_cases.parquet
---

# Synthetic Clinical Telemetry for Early-Warning Scoring

`FluxFae/Synthetic_Clinical_Telemetry` · dataset version `1.0.0` · generated `2026-07-01T00:00:00Z`

A small, **fully synthetic** telemetry dataset for practising early-warning scoring (NEWS2) on
wearable patient vitals. Every row is produced by a deterministic generator (`seed=42`): no real
patients, no PHI, no scraped data, no value sampled from a clinical source.

## What this dataset is not

- **No real patients.** No record here describes a real person.
- **Not for clinical use.** Nothing here is a medical device, a validated risk model, or clinical
  guidance. Do not deploy any model trained on it.
- **Not an annotation set.** The labels are computed from the generator's own scoring rules, so a
  model trained here learns those rules, not clinical judgement.

## Tables

| table | rows | columns | file |
| --- | --- | --- | --- |
| `patients` | 500 | 5 | `patients.csv` |
| `telemetry_vitals` | 6000 | 22 | `telemetry_vitals.csv` |
| `telemetry_edge_cases` | 50 | 11 | `telemetry_edge_cases.csv` |

Each table is published twice: the Parquet file at the repository root (what the viewer and
`load_dataset` read) and the generator's original artifact under `data/`, byte-identical to the
closed export. Both carry the same values.

`patients` holds one row per patient and a `split` column with the frozen `train` / `validation` /
`test` assignment (assigned per patient, so no patient crosses splits). `telemetry_vitals` holds the
readings, joined to a patient by `patient_id`; the `split` of a reading is the split of its patient.
`telemetry_edge_cases` holds deliberately malformed readings — out-of-range values and one row with
every vital missing — for a validator to catch; it is a test fixture, not training data.

## How to load it

```python
from datasets import load_dataset

vitals = load_dataset("FluxFae/Synthetic_Clinical_Telemetry", "telemetry_vitals", split="train")
print(vitals.features)
print(vitals[0])

patients = load_dataset("FluxFae/Synthetic_Clinical_Telemetry", "patients", split="train")
edge_cases = load_dataset("FluxFae/Synthetic_Clinical_Telemetry", "telemetry_edge_cases", split="train")
```

Every config declares a single `train` split; the logical `train` / `validation` / `test`
assignment travels as data in the `split` column, so a resampling step of your own stays explicit:

```python
train = patients.filter(lambda row: row["split"] == "train")
```

Raw artifacts, for consumers that want the files exactly as exported:

```python
from huggingface_hub import hf_hub_download

patients_csv = hf_hub_download("FluxFae/Synthetic_Clinical_Telemetry", "data/patients.csv", repo_type="dataset")
```

## Read the dictionary first

`data/data_dictionary.md` documents every column: type, role (`id`, `time`, `feature`, `target`,
`protected`, `metadata`, `split`), range, allowed values and derivation. Read it before using any
column. Several columns are targets (`news2_score_current`, `future_max_score_6h`,
`deterioration_risk_6h`, `critical_risk_6h`) and must never be used as model inputs, and at least one
column is a protected attribute intended only for fairness checks.

## Provenance

| field | value |
| --- | --- |
| dataset version | `1.0.0` |
| seed | `42` |
| generation timestamp | `2026-07-01T00:00:00Z` (fixed, never the wall clock) |
| determinism | two independent runs produce byte-identical files |

`release_manifest.json` at the repository root records, for every file this release publishes, its
size and `sha256`, plus the `sha256` of the source export's own manifest. Verify a download against
it before trusting a value.

This repository publishes the **telemetry tables only**. The v1.0.0 export also contains a
500-row operator-ticket corpus; it is not part of this dataset, so `data/validation_report.json`
describes the full export, including the two checks that mention that ticket table. Nothing else in
the report is affected: the fourteen checks describe the telemetry data published here, and all of
them pass.

## Intended use

- Practising early-warning-score pipelines and deterioration-label logic on tabular telemetry.
- Benchmarking data validators, schema checks and reproducibility tooling against the edge-case
  table.
- Teaching time-series feature engineering under a strict no-leakage contract (label windows are
  declared per row).

Out of scope: clinical decision support, publication of clinical findings, any claim about real
patient populations, and any fairness conclusion about real people.

## License

Released under [CC BY 4.0](https://creativecommons.org/licenses/by/4.0/). You may use, share and
adapt the data, including commercially, as long as you give appropriate credit.

## Citation

```bibtex
@dataset{creator_dataset_telemetry_v1.0.0,
  title  = {Synthetic Clinical Telemetry for Early-Warning Scoring},
  year   = {2026},
  note   = {Synthetic dataset, CC BY 4.0; telemetry subset of Creator Dataset v1.0.0},
  url    = {https://huggingface.co/datasets/FluxFae/Synthetic_Clinical_Telemetry}
}
```