README.md

---
license: cc-by-nc-4.0
task_categories:
  - tabular-classification
  - tabular-regression
language:
  - en
tags:
  - synthetic
  - epilepsy
  - seizures
  - neurology
  - ilae
  - sanad
  - aed
  - drug-resistant-epilepsy
  - eeg
  - ied
  - hfo
  - qolie-31
  - sudep
  - vagus-nerve-stimulation
  - ketogenic-diet
  - epilepsy-surgery
  - hippocampal-sclerosis
  - photosensitive-epilepsy
  - levetiracetam
  - lamotrigine
  - valproate
pretty_name: "HC-NEU-003 — Epilepsy Seizure Event Dataset (Sample)"
size_categories:
  - 1K<n<10K
---

# HC-NEU-003 — Epilepsy Seizure Event Dataset (Sample)

A schema-identical preview of **HC-NEU-003**, the XpertSystems.ai synthetic
**epilepsy patient cohort** dataset for ILAE 2017 seizure classification
research, antiepileptic drug (AED) efficacy modeling, EEG biomarker
analysis, SUDEP risk stratification, and seizure prediction machine
learning. The full product covers 10,000 patients with 5-year follow-up
summaries. This sample is HF-sized at 3,000 patients (one row per patient).

> **Built by** XpertSystems.ai — Synthetic Data Platform
> **Contact** [pradeep@xpertsystems.ai](mailto:pradeep@xpertsystems.ai) · [xpertsystems.ai](https://xpertsystems.ai)
> **License** CC-BY-NC-4.0 (sample); commercial license available for the full product.

---

## What HC-NEU-003 does — and how it positions the Healthcare/Neurology vertical

HC-NEU-003 is the **third Healthcare / Neurology SKU** in the XpertSystems
catalog (HC-NEU-001 = Alzheimer's, HC-NEU-002 = Parkinson's). Together,
these three SKUs cover the **three most heavily studied neurological
diseases by patient count and pharma R&D spend**:

| SKU | Disease | US Patients | Pharma R&D | Cadence |
|---|---|---|---|---|
| HC-NEU-001 | Alzheimer's | ~6.9M | ~$8B | Longitudinal, 20 visits, semi-annual |
| HC-NEU-002 | Parkinson's | ~1.0M | ~$5B | Longitudinal, 32 visits, quarterly |
| HC-NEU-003 | **Epilepsy** | **~3.4M** | **~$3B** | **Cross-sectional, 5-year summary** |

Epilepsy is a fundamentally different disease shape than AD/PD — most
clinical decisions hinge on **5-year medication trials and event counts**
rather than continuous decline. HC-NEU-003 reflects this by shipping a
**cross-sectional dataset** (one row per patient, ~89 columns) where each
row summarizes a 5-year follow-up window: seizure characterization,
trigger patterns, EEG findings, AED treatment + response, neuroimaging,
quality of life, and SUDEP risk.

This is the substrate **epilepsy pharma teams, neuro-stimulation
device vendors, EEG analytics platforms, seizure prediction ML
researchers, and ILAE epidemiology teams** have been waiting for: a
coherent dataset where seizure type × etiology × AED × EEG biomarkers ×
QoL × SUDEP risk all interact with **SANAD-trial-grade calibration**.

| Buyer Persona | Use Case |
|---|---|
| Epilepsy Pharma R&D | AED efficacy comparator modeling, DRE patient stratification |
| Neuro-Stim Device Vendor | VNS/RNS/DBS candidate identification ML |
| EEG Analytics Platform | IED detection ML, ictal pattern classification |
| Seizure Prediction ML | Trigger-pattern + sensor + EEG ML training |
| ILAE Epidemiology Team | Population-level epilepsy statistics modeling |
| SUDEP Risk Modeling | Multi-factor SUDEP-7 risk stratification |
| Surgical Candidate Selection | MRI + PET + EEG concordance ML for epilepsy surgery |

---

## What's inside

**Single cross-sectional dataframe**, one row per patient.

| Output | Rows (sample) | Columns | Size |
|---|---:|---:|---|
| `HC_NEU_003_dataset.csv` | 3,000 | 89 | ~1.4 MB |

Schema is provided in `HC_NEU_003_schema.json`.

The schema spans 7 modality blocks (demographics + seizure events +
triggers + EEG + treatment + neuroimaging + QoL/comorbidities) covering
ILAE-aligned seizure characterization, AED pharmacology, EEG biomarkers,
and patient-reported outcomes.

---

## Calibration sources

Every distribution is anchored to **named clinical references**. The
headline anchors are **SANAD I** (Marson 2007 Lancet) and **SANAD II**
(Marson 2021 Lancet) for AED-specific seizure-free rates, **Kwan & Brodie
2010** ILAE Task Force for drug-resistant epilepsy, and **WHO GBD Epilepsy
Atlas** for etiology distribution. Other anchors:

- **SANAD I trial (Marson 2007 Lancet)** — Levetiracetam vs Lamotrigine
  vs Carbamazepine in focal epilepsy.
- **SANAD II trial (Marson 2021 Lancet)** — Valproate vs Levetiracetam
  in generalised epilepsy.
- **ILAE 2017 Seizure Classification (Fisher 2017)** — 9-class seizure
  type taxonomy.
- **ILAE 2017 Etiology Framework** — 6-class etiology taxonomy.
- **Kwan & Brodie 2010 + ILAE Task Force** — drug-resistant epilepsy
  definition (failure of ≥2 appropriate AED trials).
- **WHO GBD Epilepsy Atlas** — global epilepsy prevalence and
  etiology distribution.
- **Rochester Epilepsy Study (Cramer 1998)** — QOLIE-31 normative
  data.
- **Sperling 2003 + Bernasconi 2019** — MRI lesion detection in focal
  epilepsy.
- **Fisher 2014** — photosensitive epilepsy prevalence.
- **Faught 2009 Neurology** — real-world AED adherence.
- **Devinsky 2016 + DeGiorgio 2010 SUDEP-7** — SUDEP risk
  stratification.
- **Trinka 2015 + Sperling 1990** — status epilepticus and EEG IED
  detection.

---

## Validation scorecard

The wrapper ships a 10-metric clinical-trial-anchored scorecard
(`validation_scorecard.json`) that re-scores the dataset on every
generation. Default seed 42 result:

| ID | Metric | Target | Observed | Source |
|---|---|---|---:|---|
| M01 | Levetiracetam Seizure-Free Rate | 0.27–0.43 | **0.389** | **SANAD I** |
| M02 | Lamotrigine Seizure-Free Rate | 0.36–0.52 | **0.442** | **SANAD I** |
| M03 | Valproate Seizure-Free Rate | 0.43–0.59 | **0.487** | **SANAD II** |
| M04 | Drug-Resistant Epilepsy Share | 0.26–0.46 | **0.417** | Kwan & Brodie 2010 |
| M05 | Structural Etiology Share | 0.25–0.35 | **0.306** | WHO GBD / ILAE 2017 |
| M06 | Focal Seizure Share | 0.45–0.65 | **0.532** | ILAE 2017 (Fisher 2017) |
| M07 | QOLIE-31 Score Mean | 52–68 | **55.48** | Rochester Epilepsy Study |
| M08 | Photosensitivity Share | 0.010–0.060 | **0.047** | Fisher 2014 |
| M09 | MRI Lesion Detection Share | 0.22–0.38 | **0.298** | Sperling 2003 |
| M10 | AED Adherence Mean (%) | 68–88 | **80.54** | Faught 2009 |

**Grade: A+ (100/100). Verified across seeds 42, 7, 123, 2024, 99, 1.**

**The standout achievement here is M01-M03: AED-by-AED seizure-free
rates land within 4% of their SANAD I/II trial published rates.**

- Levetiracetam observed 38.9% vs SANAD I published 35% (4 pp deviation)
- **Lamotrigine observed 44.2% vs SANAD I published 44%** (0.2 pp — exact)
- Valproate observed 48.7% vs SANAD II published 51% (2 pp deviation)

This is **SANAD-trial-grade calibration** — the synthetic data reproduces
the *exact* relative efficacy ranking of the AEDs from the two pivotal
UK NIHR-funded trials that have defined epilepsy first-line therapy
guidelines for the last 20 years.

---

## Suggested use cases

- **AED comparator modeling** — calibrated AED-specific seizure-free
  rates × adherence × side effect burden support treatment-effect
  modeling for new AED candidates.
- **Drug-resistant epilepsy (DRE) prediction** — pre-built
  `drug_resistant_epilepsy_flag` (Kwan-Brodie ILAE definition) with
  patient features enables DRE risk ML.
- **Surgical candidate identification** — `surgical_candidate_flag` +
  MRI lesion + PET hypometabolism + EEG concordance supports epilepsy
  surgery patient selection ML.
- **EEG biomarker analysis** — spectral power (delta/theta/alpha/beta/
  gamma %) + IED frequency + HFO rate + ictal pattern × seizure type
  enables EEG ML training.
- **SUDEP risk stratification** — `sudep_7_risk_score` + clinical
  features support SUDEP-7-aligned risk modeling for high-risk patient
  identification.
- **Quality of life modeling** — QOLIE-31 + NDDI-E + GAD-7 + MoCA
  multimodal QoL ML for epilepsy outcomes research.
- **Trigger pattern ML** — 9-trigger binary indicators (sleep, missed
  dose, stress, alcohol, photic, hormonal, fever, exertion, drug
  interaction) for seizure prediction model training.
- **VNS / RNS / DBS candidate selection** — for neuro-stim device vendor
  ML training on patient eligibility scoring.

---

## Loading

```python
from datasets import load_dataset

ds = load_dataset(
    "xpertsystems/hcneu003-sample",
    data_files="HC_NEU_003_dataset.csv",
    split="train",
)
```

Or with pandas directly:

```python
import pandas as pd
from huggingface_hub import hf_hub_download

path = hf_hub_download(
    repo_id="xpertsystems/hcneu003-sample",
    filename="HC_NEU_003_dataset.csv",
    repo_type="dataset",
)
df = pd.read_csv(path)
```

The dataset ships with `HC_NEU_003_schema.json` providing per-column
dtypes for pipeline integration:

```python
import json
schema = json.load(open("HC_NEU_003_schema.json"))
# {"patient_id": "object", "seizure_type": "object", "qolie_31_score": "float64", ...}
```

Unlike HC-NEU-001 (longitudinal AD) and HC-NEU-002 (longitudinal PD),
HC-NEU-003 is **cross-sectional** — one row per patient summarizing a
5-year follow-up. There are no `visit_number` or `visit_date` columns
beyond a single most-recent visit reference.

---

## Schema highlights

**Demographics** — `patient_id`, `age_at_onset_years`, `age_at_visit_years`,
`epilepsy_duration_years`, `sex`, `etiology` (6-class: Structural,
Genetic, Infectious, Metabolic, Immune, Unknown), `early_onset_flag`.

**Seizure characterization (ILAE 2017)** — `seizure_type` (9-class:
Focal_Aware, Focal_Impaired_Awareness, Focal_to_Bilateral_Tonic_Clonic,
Absence, Myoclonic, Tonic_Clonic, Atonic, Tonic, Epileptic_Spasm),
`seizure_onset_zone`, `seizure_frequency_per_month`, `seizure_duration_sec`,
`postictal_duration_min`, `status_epilepticus_flag`, `seizure_cluster_flag`,
`nocturnal_flag`, `provoked_flag`, `annualized_seizure_rate`,
`seizure_free_days`, `seizure_diary_compliance_pct`,
`seizure_semiology_code`.

**Triggers** — `trigger_sleep_deprivation`, `trigger_missed_aed_dose`,
`trigger_emotional_stress`, `trigger_alcohol_use`, `trigger_photosensitivity`,
`trigger_hormonal`, `trigger_fever_infection`, `trigger_physical_exertion`,
`trigger_medication_interaction`, `trigger_count`,
`trigger_identified_flag`, `catamenial_pattern_flag`.

**EEG** — `eeg_type` ∈ {Scalp_Routine, Scalp_Ambulatory, sEEG, ECoG},
`eeg_duration_hours`, `eeg_background_activity`,
`eeg_interictal_discharge_flag`, `eeg_ied_morphology`,
`eeg_ied_frequency_per_hour`, `eeg_ictal_pattern`,
`eeg_ictal_onset_frequency_hz`, spectral power % (delta/theta/alpha/beta/
gamma), `eeg_hfo_rate_per_min`, `eeg_spike_wave_hz`,
`eeg_seizure_captured_flag`.

**Treatment** — `treatment_arm` ∈ {Monotherapy, Polytherapy, Surgical,
VNS_Therapy, Ketogenic_Diet, No_AED}, `aed_name` (13 agents),
`aed_dose_mg_day`, `aed_mechanism`, `aed_polytherapy_count`,
`aed_duration_months`, `aed_adherence_pct`, `seizure_reduction_pct`,
`seizure_free_flag`, `treatment_response_class` ∈ {Seizure_Free,
Responder, Non_Responder}, `drug_resistant_epilepsy_flag`,
`aed_side_effect_score` (SIDAED 0-20), `aed_serum_level_ug_ml`,
`surgical_candidate_flag`, `vns_therapy_flag`, `ketogenic_diet_flag`.

**Neuroimaging** — `mri_lesion_type` ∈ {Hippocampal_Sclerosis, FCD,
Vascular, Tumor, Cortical_Malformation, NaN}, `mri_lesion_flag`,
`hippocampal_volume_ml`, `hippocampal_asymmetry_index`, `fcd_location`,
`pet_hypometabolism_flag`, `spect_lateralization_index`,
`mri_field_strength_T`.

**QoL & comorbidities** — `qolie_31_score` (0-100), `qolie_seizure_worry`,
`nddi_e_score` (0-30), `nddi_e_positive_flag`, `anxiety_gad7_score`
(0-21), `cognitive_impairment_flag`, `moca_score`,
`driving_restriction_flag`, `employment_status` ∈ {Employed, Unemployed,
Disability, Student, Retired}, `sudep_7_risk_score`, `sudep_high_risk_flag`,
`psychiatric_comorbidity` ∈ {Depression, Anxiety, PNES, ADHD, NaN},
`charlson_comorbidity_index`, `bmi`.

---

## Calibration notes & limitations

In the spirit of honest synthetic data, a few things buyers of the sample
should know:

1. **Cross-sectional, not longitudinal.** Unlike HC-NEU-001 (20 visits)
   and HC-NEU-002 (32 visits), HC-NEU-003 is one-row-per-patient
   summarizing a 5-year follow-up. This is the **clinically appropriate
   shape for epilepsy management** (medication trials measured in years,
   not weeks), but means time-series modeling on this sample requires
   different methodology than HC-NEU-001/002.

2. **EEG IED detection share ~68% — below clinical 80-90%.** Salinsky
   1987 reports IED detection in 50% of single routine EEGs but 80-90%
   with repeat EEG or sleep-deprived EEG. The generator's `eeg_type` mix
   is dominated by routine EEG (49%), reflecting realistic clinical
   acquisition patterns. For higher-yield IED studies, filter by
   `eeg_type` ∈ {Scalp_Ambulatory, sEEG} to recover ≥85% IED detection.

3. **Status epilepticus rate ~2.4% — below epilepsy lifetime 5-10%.**
   Trinka 2015 reports lifetime SE risk 5-10% in epilepsy populations;
   the generator reports per-patient SE history rate ~2.4%. The full
   product calibrates SE history more aggressively for lifetime risk
   modeling.

4. **SUDEP high-risk flag ~2.5% — at low end of high-risk cohort range
   5-10%.** The actual SUDEP incidence is ~1/1000 patient-years
   (Devinsky 2016); cohorts classified "high SUDEP risk" by SUDEP-7
   typically span 5-10% of populations. The generator is conservative;
   use `sudep_7_risk_score > 75th percentile` for relative risk
   stratification.

5. **Sleep deprivation trigger ~29% — at low end of Sperling 30-40%.**
   Patient self-report varies widely; the generator's 29% rate reflects
   conservative trigger attribution.

6. **Catamenial pattern flag ~5.7% — well below Herzog 1997 30% of
   female PD patients.** This reflects the *strict definition*
   (perimenstrual + periovulatory + perimenstrual+anovulatory subtypes
   combined); broader self-reported catamenial epilepsy is higher.
   For broader catamenial filtering, intersect with
   `trigger_hormonal == 1`.

7. **Driving restriction flag 89% — reflects DRE-enriched cohort.**
   This sample includes 42% DRE patients (above general population
   30-40%); driving restriction follows DMV/state-by-state rules. For
   non-DRE driving analysis, filter `drug_resistant_epilepsy_flag == 0`.

8. **Treatment arm "No_AED" subset (~4.5%)** represents seizure-free
   recently-tapered, single-unprovoked-seizure not yet started AED, and
   refractory-no-current-AED subgroups combined. For deployed-patient
   modeling, filter to `treatment_arm != "No_AED"`.

9. **Cross-sectional means no within-patient trajectory.** For
   medication-switching modeling (a common clinical scenario), use the
   full product which carries `aed_history` as a list-typed column with
   per-AED entries.

10. **Deterministic seeding.** Wrapper passes user-specified seed into
    `CONFIG["seed"]`, `np.random.seed()`, and `random.seed()`. Seed
    sweep verifies Grade A+ across {42, 7, 123, 2024, 99, 1}.

---

## Commercial / full product

The full **HC-NEU-003** product covers 10,000 patients × 5-year follow-up
with optional longitudinal sub-cohort (per-quarter seizure event diaries),
AED-switching history with sequence modeling, refined SUDEP risk modeling
including detailed mortality cohort, ethnic-population stratified AED
metabolism modeling (CYP-genotype-stratified phenytoin/carbamazepine
levels), and pediatric epilepsy cohort variants (Lennox-Gastaut, Dravet,
West syndrome). Available under commercial license — contact
[pradeep@xpertsystems.ai](mailto:pradeep@xpertsystems.ai).

XpertSystems.ai also publishes synthetic data products across Oil & Gas
(17 SKUs), Cybersecurity, Insurance & Risk, and Materials & Energy.
Catalog: [huggingface.co/xpertsystems](https://huggingface.co/xpertsystems).