electricsheepafrica/brain-tumor-single-slice-MRI-scan-with-synthetic-ehr-africa

Dataset Card: Africa Brain Tumor Scans with Synthetic EHR (Bundled Parquet)

This dataset bundles single-slice brain MRI scans and richly structured, synthetic EHR data into a single Parquet file suitable for multimodal ML research. Each row contains an image struct (bytes + path), a source label column, and an EHR payload with both a full JSON record and convenient summary columns.

The synthetic EHRs are Africa-focused: they encode country, urban/rural, facility level, insurance, time-to-presentation, common comorbidities (e.g., HIV/TB/malaria), realistic procedures (CPT/SNOMED/ICD-10-PCS), diagnoses (ICD-10), medications with dosing and ATC codes, longitudinal encounters, and outcomes timelines. Clinical rules and a strict validator enforce core coherence (e.g., temozolomide only in tumor cases; AEDs when seizures are present; longer delays for rural/primary settings).

All data are synthetic and non-identifiable.

Size and Composition

• Rows: 7,023
• File size: ~165.4 MB (165,397,598 bytes)
• Labels (from source label column):
  • 0: 1,621
  • 1: 1,645
  • 2: 2,000
  • 3: 1,757

Note on labels: The EHR generator interprets labels > 0 as “tumor-present” for clinical logic. The ehr_json includes a tumor_present field mirroring the source label integer. If you need a strict binary, derive tumor_binary = 1 if label > 0 else 0 (see example below).

Demographics and Clinical Context (Aggregates)

From outputs/ehr/brain_tumor_ehr.csv (flattened EHRs):

• Gender: 3,618 M; 3,405 F
• Age: min 25, max 85, mean 55.1, median 55.0
• Country (top 10): Nigeria 736; Egypt 726; Morocco 720; Uganda 706; Tanzania 700; Kenya 698; Ghana 698; South Africa 687; Ethiopia 679; Rwanda 673
• Facility level: primary 2,929; secondary 2,323; tertiary 1,771
• Urban/rural: urban 4,349; rural 2,674
• Triage acuity: routine 3,323; urgent 2,842; emergency 858
• Provider specialty: neurology 2,794; neurosurgery 2,677; internal_medicine 1,552
• Time to presentation (days): urban mean 10.64; rural mean 26.27

Coherence examples (sanity checks):

• Antiepileptics when “seizures” present: 100.0% vs 10.8% without seizures
• Temozolomide in tumor vs non-tumor: 44.0% vs 0.0%
• Dexamethasone in tumor vs non-tumor: 89.7% vs 0.0%

Access and Usage

Python (PyArrow + Pandas):

import pyarrow.parquet as pq
import json

path = "outputs/bundled/africa_brain_tumor_scans_with_ehr.parquet"
tbl = pq.read_table(path)
df = tbl.to_pandas()

# Parse the EHR JSON for first row
ehr = json.loads(df.loc[0, "ehr_json"])  # dict
print(ehr["diagnoses"][0])

# Derive binary tumor label if needed
df["tumor_binary"] = (df["label"].astype(int) > 0).astype(int)

Image decoding:

• image.bytes holds the encoded image bytes; image.path is the original path string (may be local-only).
• Use src/parquet_image_io.py helpers (e.g., decode_to_pil) or Pillow directly to load images from bytes.

CSV/Parquet EHR (flattened):

• JSON columns (e.g., diagnoses, procedures) are JSON strings; use json.loads to parse.
• Lists of primitives (e.g., symptoms) are semicolon-joined strings.

Ethical Considerations and Limitations

• Synthetic, non-identifiable data; no PHI.
• Distributions are plausibility-driven, not institution-specific; do not perform epidemiologic inference.
• Africa profile encodes broad tendencies (e.g., delays, comorbidities) and standard code systems (ICD-10, SNOMED, CPT/ICD-10-PCS); mappings are illustrative.
• Generator is rule-driven and may not capture all real-world correlations.

License and Contact

• See repository LICENSE for terms (if absent, treat as “all rights reserved” by default).
• Questions or calibration requests: open an issue in the repository.