hlt009-sample / README.md
pradeep-xpert's picture
Upload folder using huggingface_hub
0d08ba6 verified
---
license: cc-by-nc-4.0
task_categories:
- tabular-classification
- tabular-regression
- time-series-forecasting
language:
- en
tags:
- synthetic
- healthcare
- icu
- vital-signs
- continuous-monitoring
- mimic-iv
- eicu-crd
- time-series
- apache-ii
- sofa
- news2
- qsofa
- charlson
- alarm-fatigue
- joint-commission
- iec-60601
- ventilation
- vasopressor
- sepsis
- early-warning-score
- deterioration-prediction
- signal-quality
- monitoring-devices
- philips-intellivue
- ge-carescape
- masimo
- nihon-kohden
pretty_name: HLT-009 Synthetic Continuous Vital Sign Monitoring Dataset (Sample Preview)
size_categories:
- 10K<n<100K
---
# HLT-009 — Synthetic Continuous Vital Sign Monitoring Dataset (Sample Preview)
**A free, schema-identical preview of the full HLT-009 commercial product from [XpertSystems.ai](https://xpertsystems.ai).**
A **fully synthetic** ICU continuous vital sign monitoring dataset combining 12-stream time-series vitals (HR/SpO2/RR/NBP/IBP/Temp/EtCO2/CVP/CO), alarm event logs with true/false labels, intervention logs (medication boluses, ventilator adjustments, code events), and 53-column episode-level summary data — calibrated to MIMIC-IV / eICU-CRD benchmarks with APACHE-II, SOFA, NEWS2, qSOFA, and CCI severity scoring.
> ⚠️ **PRIVACY & SYNTHETIC NATURE**
> Every record in this dataset is **100% synthetic**. **No real patient data, no PHI, no real medical device readings.** Population-level distributions match published MIMIC-IV / eICU-CRD / Drew et al. benchmarks but the episodes and waveforms are computationally generated.
---
## What's in this sample
| File | Rows | Cols | Description |
|---|---|---|---|
| `vitals_timeseries.csv` | ~26,700 | 19 | One row per episode-timestep (5-min resolution). 12 vital streams + NEWS2 + qSOFA + artifact flag + 6h rolling features |
| `alarm_events.csv` | ~170 | 15 | One row per alarm event. Type, priority (IEC 60601-1-8), true/false flag, false-alarm subtype, response time, override flag |
| `interventions.csv` | ~15 | 8 | One row per clinical intervention (medication bolus, ventilator adjustment, code event, rapid response) |
| `episode_summary.csv` | 25 | 53 | One row per episode. Demographics, APACHE-II, SOFA, CCI, ventilation/vasopressor/RRT flags, LOS, NEWS2 max/mean, deterioration label, mortality, 12 signal quality indices |
**Total:** ~5.3 MB across 5 files.
---
## Schema highlights
### `vitals_timeseries.csv` (19 columns, ~1,100 rows per episode at 5-min resolution)
**Identity:** `episode_id`, `timestamp`
**12 vital streams** (calibrated to MIMIC-IV physiological ranges):
- **Cardiovascular:** `hr_bpm`, `nbp_sys_mmhg`, `nbp_dia_mmhg`, `nbp_map_mmhg`, `ibp_sys_mmhg`, `ibp_dia_mmhg`, `cvp_mmhg`, `cardiac_output_lpm`
- **Respiratory:** `spo2_pct`, `rr_bpm`, `etco2_mmhg`
- **Thermoregulation:** `temp_c`
**Derived & quality:** `artifact_flag` (4% rate per timestep), `news2_score` (RCP NEWS2 computed at each step), `qsofa_score` (Sepsis-3 qSOFA), `news2_roll_max_4h`, `news2_rate_of_rise`
### `alarm_events.csv` (15 columns)
`alarm_id`, `episode_id`, `alarm_type` (18 types: HIGH_HR, LOW_HR, CRITICAL_LOW_HR, LOW_SPO2, CRITICAL_LOW_SPO2, HIGH_RR, APNEA, HIGH_SBP, LOW_SBP, LOW_MAP, HIGH_ETCO2, LOW_ETCO2, HIGH_CVP, LOW_CVP, HIGH_CO, LOW_CO, HIGH_IBP_SYS, LOW_IBP_SYS), `alarm_priority` (IEC 60601-1-8: LOW/MEDIUM/HIGH/CRITICAL), `alarm_onset_ts`, `alarm_duration_sec`, `true_alarm_flag`, `false_alarm_subtype` (Artifact / Motion / LeadOff / TechnicalError), `response_time_min`, `intervention_triggered`, `override_flag`, `limit_at_alarm_low`, `limit_at_alarm_high`, `alarm_cascade_id`, `shift` (Day/Evening/Night)
### `interventions.csv` (8 columns)
`intervention_id`, `episode_id`, `intervention_type` (MEDICATION_BOLUS / VENTILATOR_ADJUSTMENT / POSITION_CHANGE / PHYSICIAN_NOTIFICATION / RAPID_RESPONSE_ACTIVATION / CODE_EVENT / NURSING_ASSESSMENT), `intervention_ts`, `triggered_by_alarm`, `time_from_alarm_min`, `clinician_role`, `intervention_outcome`
### `episode_summary.csv` (53 columns)
**Identity & setting:** `episode_id`, `monitoring_setting` (ICU), `icu_unit_type` (MICU/SICU/CCU/Neuro ICU), `bed_id`, `admit_dt`, `discharge_dt`, `episode_duration_days`
**Demographics & severity:** `age`, `sex`, `apache2_score` (Knaus 1985), `sofa_score` (Vincent 1996), `sofa_at_discharge`, `cci_score` (Charlson 1987), `primary_dx_group` (Sepsis/Respiratory Failure/Cardiac/Neuro/Post-Surgical/Trauma/Other), `trajectory` (Stable/Improving/Deteriorating/Oscillating)
**Clinical interventions:** `ventilation_status`, `vasopressor_flag`, `rrt_flag` (renal replacement therapy), `has_arterial_line`, `has_central_line`, `has_pa_catheter`
**Device metadata:** `monitor_manufacturer` (Philips IntelliVue MX800 / GE Carescape B850 / Masimo Root / Nihon Kohden BSM-6000), `rpm_device_type`, `lead_configuration` (3-lead / 5-lead / 12-lead), `device_uptime_pct`, `connectivity_drops`
**Alarm fatigue metrics (Drew et al. 2014):** `true_alarm_rate`, `total_alarms`, `alarms_per_patient_day`, `actionable_alarm_rate`, `alarm_override_rate`, `median_response_time_min`, `alarm_limit_modification_count`, `alarm_cascade_count`, `fatigue_index_score`
**Early warning & outcomes:** `news2_max`, `news2_mean`, `qsofa_max`, `deterioration_6h_label`, `in_hospital_mortality`, `readmission_30d`, `rapid_response_event`
**Signal Quality Indices (SQI):** 12 columns `sqi_*` — one per vital stream
---
## Calibration source story
The full HLT-009 generator anchors all distributions to authoritative critical care references:
- **MIMIC-IV (Johnson et al. Scientific Data 2023)** — ICU vital signs benchmark, LOS Weibull(1.4, 5.2), severity distributions
- **eICU-CRD (Pollard et al. Scientific Data 2018)** — Multi-center ICU database, ventilation/vasopressor rates
- **APACHE-II (Knaus et al. Crit Care Med 1985)** — Acute Physiology and Chronic Health Evaluation
- **SOFA (Vincent et al. Intensive Care Med 1996)** — Sequential Organ Failure Assessment
- **NEWS2 (Royal College of Physicians 2017)** — National Early Warning Score 2
- **qSOFA (Singer et al. JAMA 2016)** — Sepsis-3 Quick SOFA
- **CCI (Charlson et al. J Chron Dis 1987)** — Charlson Comorbidity Index
- **Drew et al. (2014) PLoS ONE** — Alarm fatigue benchmark (187 alarms/bed/day)
- **Joint Commission Sentinel Event Alert 50 (2013)** — Alarm safety standards
- **Wunsch et al. (2010) JAMA** — US ICU mechanical ventilation prevalence
- **IEC 60601-1-8** — Medical electrical equipment alarm priority standard
### Sample-scale validation scorecard
| Metric | Observed | Target | Tolerance | Status | Source |
|---|---|---|---|---|---|
| Mean APACHE-II score | 11.6 | 12.0 | ±4.0 | ✅ PASS | Knaus et al. (1985) / MIMIC-IV |
| Mean SOFA score | 3.1 | 3.5 | ±2.0 | ✅ PASS | Vincent et al. (1996) |
| Median LOS (days) | 2.99 | 4.0 | ±2.0 | ✅ PASS | MIMIC-IV (Johnson et al. 2023) |
| Ventilation rate | 56% | 40% | ±20% | ✅ PASS | Wunsch et al. (2010) |
| Mean NEWS2 score | 4.29 | 4.0 | ±1.5 | ✅ PASS | RCP NEWS2 (2017) |
| True alarm rate | 17.1% | 20% | ±10% | ✅ PASS | Joint Commission SE Alert 50 |
| Artifact flag rate | 3.85% | 4% | ±2% | ✅ PASS | Wong et al. (2018) ICU data quality |
| Vital stream count | 12 | 12 | — | ✅ PASS | Schema coverage |
| Alarm priority diversity | 2 | ≥2 | — | ✅ PASS | IEC 60601-1-8 |
| Timeseries temporal monotonicity | 100% | 100% | — | ✅ PASS | Data hygiene |
**Grade: A+ (100/100) — verified across 6 random seeds (42, 7, 123, 2024, 99, 1).**
---
## Loading examples
### Pandas — explore the episode summary
```python
import pandas as pd
summary = pd.read_csv("episode_summary.csv")
vitals = pd.read_csv("vitals_timeseries.csv", parse_dates=["timestamp"])
alarms = pd.read_csv("alarm_events.csv", parse_dates=["alarm_onset_ts"])
# Severity by primary diagnosis
print(summary.groupby("primary_dx_group")[
["apache2_score", "sofa_score", "episode_duration_days"]
].mean().round(2))
# Alarm volume by ICU unit
print(summary.groupby("icu_unit_type")["alarms_per_patient_day"].mean())
```
### Hugging Face Datasets
```python
from datasets import load_dataset
ds = load_dataset("xpertsystems/hlt009-sample", data_files={
"vitals": "vitals_timeseries.csv",
"alarms": "alarm_events.csv",
"interventions": "interventions.csv",
"summary": "episode_summary.csv",
})
print(ds)
```
### Vital sign trajectory plot
```python
import pandas as pd
import matplotlib.pyplot as plt
vitals = pd.read_csv("vitals_timeseries.csv", parse_dates=["timestamp"])
# Plot HR + SpO2 trajectory for one episode
ep_id = vitals["episode_id"].iloc[0]
ep = vitals[vitals["episode_id"] == ep_id].sort_values("timestamp")
fig, (ax1, ax2) = plt.subplots(2, 1, figsize=(10, 6), sharex=True)
ax1.plot(ep["timestamp"], ep["hr_bpm"], color="#c44")
ax1.set_ylabel("HR (bpm)")
ax1.axhline(120, ls="--", color="grey", alpha=0.5) # HIGH_HR threshold
ax2.plot(ep["timestamp"], ep["spo2_pct"], color="#4488ff")
ax2.set_ylabel("SpO2 (%)")
ax2.axhline(90, ls="--", color="grey", alpha=0.5) # LOW_SPO2 threshold
ax2.set_xlabel("Time")
plt.suptitle(f"Vitals for episode {ep_id}")
plt.show()
```
### Deterioration prediction baseline
```python
import pandas as pd
import numpy as np
from sklearn.ensemble import GradientBoostingClassifier
from sklearn.model_selection import train_test_split
vitals = pd.read_csv("vitals_timeseries.csv", parse_dates=["timestamp"])
summary = pd.read_csv("episode_summary.csv")
# Build a feature matrix at episode level from first-4h vitals
first_4h_features = []
for ep_id, ep in vitals.groupby("episode_id"):
ep_sorted = ep.sort_values("timestamp")
# Use first 48 timesteps = first 4 hours at 5-min resolution
first_4h = ep_sorted.head(48)
if len(first_4h) >= 12:
first_4h_features.append({
"episode_id": ep_id,
"hr_mean": first_4h["hr_bpm"].mean(),
"hr_std": first_4h["hr_bpm"].std(),
"spo2_min": first_4h["spo2_pct"].min(),
"rr_max": first_4h["rr_bpm"].max(),
"news2_max_first4h": first_4h["news2_score"].max(),
"news2_mean_first4h": first_4h["news2_score"].mean(),
})
feats = pd.DataFrame(first_4h_features).merge(
summary[["episode_id", "apache2_score", "sofa_score", "cci_score",
"ventilation_status", "deterioration_6h_label"]],
on="episode_id"
)
X = feats.drop(["episode_id", "deterioration_6h_label"], axis=1).fillna(0)
y = feats["deterioration_6h_label"]
if y.nunique() > 1:
Xtr, Xte, ytr, yte = train_test_split(X, y, test_size=0.3, random_state=42)
m = GradientBoostingClassifier(random_state=42).fit(Xtr, ytr)
print(f"6h deterioration AUC: {m.score(Xte, yte):.3f}")
```
### Alarm fatigue analysis
```python
import pandas as pd
summary = pd.read_csv("episode_summary.csv")
alarms = pd.read_csv("alarm_events.csv")
# Fatigue index by trajectory
print(summary.groupby("trajectory")[
["alarms_per_patient_day", "true_alarm_rate", "alarm_override_rate",
"fatigue_index_score"]
].mean().round(3))
# False alarm subtypes
print(alarms[alarms["true_alarm_flag"] == 0]["false_alarm_subtype"]
.value_counts())
```
---
## Suggested use cases
- **6-hour deterioration prediction** — predict `deterioration_6h_label` from first-N-hour vitals + summary features
- **Alarm fatigue research** — analyze actionable vs nuisance alarm patterns, build false-alarm classifiers
- **Sepsis prediction** — train models on vital trajectories + qSOFA + NEWS2 trends
- **ICU mortality risk** — predict `in_hospital_mortality` from baseline severity + early vital features
- **Mechanical ventilation prediction** — predict ventilation onset from vital trajectories
- **NEWS2 / qSOFA validation** — test calibration of early warning scores in ML-augmented pipelines
- **Signal quality / artifact classification** — train artifact detectors using `sqi_*` and `artifact_flag` labels
- **Time-series anomaly detection** — vital sign outlier detection, change-point detection
- **Multi-stream time-series modeling** — joint LSTM/Transformer modeling on 12 simultaneous vital streams
- **Alarm cascade analysis** — alarm propagation and crash-cart event prediction
- **Healthcare AI MLOps** — pipeline testing for streaming ICU data, real-time inference rehearsal
- **Educational use in critical care medicine and biomedical engineering**
---
## Sample vs. full product
| Aspect | This sample | Full HLT-009 product |
|---|---|---|
| Episodes | 25 | 10,000+ (default) up to 1M |
| Settings | ICU only | mixed (ICU + RPM) configurable |
| Time resolution | 5-min | 1-min or 5-min |
| Schema | identical | identical |
| Calibration | identical | identical |
| License | CC-BY-NC-4.0 | Commercial license |
The full product unlocks:
- **Up to 1M episodes** for production-grade deterioration / sepsis / alarm fatigue model training
- **RPM (Remote Patient Monitoring)** episodes with multi-week outpatient monitoring (7-91 days)
- **1-min resolution** for high-frequency analysis
- **Mixed ICU+RPM** for cross-care-setting model training
- Commercial use rights
**Contact us for the full product.**
---
## Limitations & honest disclosures
- **Sample is preview-only.** 25 episodes × ~27K timesteps is enough to demonstrate schema and calibration, but is **not statistically sufficient** for training deterioration prediction or sepsis classifiers. Use the full product (10K+ episodes) for serious ML work.
- **ICU-only in this sample, not mixed setting.** RPM episodes average 7-91 days × 288 timesteps/day = ~14K rows each, which would push the sample past 20 MB. The full product supports mixed ICU + RPM cohorts.
- **Sample is on the larger side (5.3 MB)** because continuous vital sign data has natural fan-out — each multi-day ICU episode produces ~1,000-3,000 timesteps at 5-min resolution. The full product scales linearly with episode count.
- **Alarm priority diversity limited at this sample scale.** The schema supports 4 priority levels (LOW/MEDIUM/HIGH/CRITICAL per IEC 60601-1-8), but at n=25 only MEDIUM+HIGH alarm types fire reliably. CRITICAL alarms (CRITICAL_LOW_HR, APNEA, CRITICAL_LOW_SPO2) require extreme physiology that's rare in stable cohorts. LOW priority alarms (HIGH_CVP, HIGH_CO) are also rare. The full product produces all 4 levels at scale.
- **Vital signs are simulated, not real waveform data.** Each timestep value is sampled from physiologically-realistic distributions calibrated to MIMIC-IV ranges. This is appropriate for ML algorithm development, but does NOT capture the full beat-to-beat waveform variability of real continuous monitoring. Real waveforms exhibit autocorrelation, R-R interval variability, and respiratory modulation that this synthetic data does not fully reproduce.
- **5-minute resolution, not beat-to-beat.** The full product supports 1-min resolution; production ICU monitors sample at 125-500 Hz (waveform-level). For HRV / arrhythmia / respiratory waveform analysis, real waveform data is required.
- **Mortality rate runs slightly low at this sample size (4-16% vs MIMIC-IV target 8-15%).** At n=25 episodes this is 1-4 deaths total, so seed-to-seed variance is high. The full product hits 10-12% mortality reliably.
- **Ventilation rate runs slightly high (~50% vs target 30-45%).** This is a generator parameter (`is_ventilated = rng.random() < 0.42`) — the actual draw varies seed-to-seed.
- **Synthetic, not derived from real ICU records.** Vital sign distributions, alarm patterns, and severity scores follow published critical care references but do NOT reflect any specific real patient cohort.
---
## Ethical use guidance
This dataset is designed for:
- ICU deterioration prediction methodology development
- Alarm fatigue research and false-alarm classifier development
- Sepsis / NEWS2 / qSOFA validation methodology
- Continuous monitoring AI pipeline testing
- Educational use in critical care medicine and biomedical informatics
- Healthcare AI pretraining for time-series clinical prediction
This dataset is **not appropriate for**:
- Making clinical decisions about real patients
- FDA-regulated AI/SaMD device training (use real de-identified clinical data)
- Real-time alarm system tuning without separate validation
- Discriminatory analyses targeting protected demographic groups
---
## Companion datasets in the Healthcare vertical
- [HLT-001](https://huggingface.co/datasets/xpertsystems/hlt001-sample) — Synthetic Patient Population (5K patients × 79 cols, CDC/NHANES calibrated)
- [HLT-002](https://huggingface.co/datasets/xpertsystems/hlt002-sample) — Synthetic EHR Dataset (4K encounters + FHIR R4 bundles)
- [HLT-003](https://huggingface.co/datasets/xpertsystems/hlt003-sample) — Synthetic Clinical Trial Dataset (3 endpoint types + power sweep)
- [HLT-004](https://huggingface.co/datasets/xpertsystems/hlt004-sample) — Synthetic Disease Progression Dataset (NSCLC + Heart Failure longitudinal)
- [HLT-005](https://huggingface.co/datasets/xpertsystems/hlt005-sample) — Synthetic Hospital Admission Dataset (5K admissions + bed utilization)
- [HLT-006](https://huggingface.co/datasets/xpertsystems/hlt006-sample) — Synthetic Medical Imaging Dataset (1K studies + COCO annotations + reports)
- [HLT-007](https://huggingface.co/datasets/xpertsystems/hlt007-sample) — Synthetic Drug Response Dataset (3K patient-treatments × 25 drug classes + PGx + PK)
- [HLT-008](https://huggingface.co/datasets/xpertsystems/hlt008-sample) — Synthetic Healthcare Claims Dataset (500 members + 30K claims + fraud labels)
- **HLT-009** — Synthetic Continuous Vital Sign Monitoring Dataset (you are here)
Use **HLT-001 through HLT-009 together** for the full healthcare ML data stack: population → EHR → trials → progression → hospital ops → imaging → pharmacology → claims → continuous monitoring.
---
## Citation
If you use this dataset, please cite:
```bibtex
@dataset{xpertsystems_hlt009_sample_2026,
author = {XpertSystems.ai},
title = {HLT-009 Synthetic Continuous Vital Sign Monitoring Dataset (Sample Preview)},
year = 2026,
publisher = {Hugging Face},
url = {https://huggingface.co/datasets/xpertsystems/hlt009-sample}
}
```
---
## Contact
- **Web:** [https://xpertsystems.ai](https://xpertsystems.ai)
- **Email:** [pradeep@xpertsystems.ai](mailto:pradeep@xpertsystems.ai)
- **Full product catalog:** Cybersecurity, Insurance & Risk, Materials & Energy, Oil & Gas, Healthcare, and more
**Sample License:** CC-BY-NC-4.0 (Creative Commons Attribution-NonCommercial 4.0)
**Full product License:** Commercial — please contact for pricing.