electricsheepafrica/Africa-skin-cancer-images-EHR

Africanized Skin Cancer Multimodal Dataset

Dataset Description

This dataset extends the HAM10000 dermatoscopic image collection with:

1. Provisional skin-tone annotations (Fitzpatrick scale estimates)
2. Synthetic African-focused Electronic Health Records (EHR)
3. Multimodal patient demographics, comorbidities, medications, and lab values

Created by: kossiso Royce
Date: October 2025
Version: 1.0

Objective

Create a more representative dataset for training and evaluating skin cancer detection models in African populations, addressing the severe underrepresentation of darker skin tones in existing dermatology datasets.

---

Dataset Composition

Images & Metadata

• Source: HAM10000 (10,015 dermatoscopic images)
• Format: JPG, 600×450 pixels
• Diagnoses: 7 classes
  • nv (melanocytic nevi): 6,705 (66.9%)
  • mel (melanoma): 1,113 (11.1%)
  • bkl (benign keratosis): 1,099 (11.0%)
  • bcc (basal cell carcinoma): 514 (5.1%)
  • akiec (actinic keratoses): 327 (3.3%)
  • vasc (vascular lesions): 142 (1.4%)
  • df (dermatofibroma): 115 (1.1%)

Skin-Tone Annotations

• Method: Automated estimation via KMeans clustering in LAB color space
• Distribution:
  • Fitzpatrick I–II (light): 8,906 images (88.9%)
  • Fitzpatrick III–IV (medium): 675 images (6.7%)
  • Fitzpatrick V–VI (dark): 434 images (4.3%)
• Status: Provisional; requires expert validation

Synthetic EHR Data

Patients (7,470 records)

• Demographics: country, region (urban/rural), occupation, sun exposure, insurance status
• Countries: Nigeria (25%), South Africa (21%), Kenya (15%), Ghana (10%), Ethiopia (10%), Tanzania (8%), Uganda (7%), Other (5%)
• Insurance: 60% uninsured, 25% public, 10% private, 5% community

Comorbidities (2,145 records)

• Hypertension: 1,052 cases (14.1% prevalence)
• HIV: 617 cases (8.3% prevalence)
• Diabetes mellitus type 2: 462 cases (6.2% prevalence)
• Albinism: 14 cases (0.2% prevalence)

Medications (3,365 records)

• Antiretrovirals (1,851), antihypertensives (1,052), antidiabetics (462)
• Linked to patient comorbidities

Lab Values (8,549 records)

• Hemoglobin (all patients), CD4 count (HIV+), fasting glucose (diabetics)

---

Data Files

data/
├── metadata.csv                    # Image metadata + skin-tone annotations
├── patients_synthetic.csv          # Patient demographics
├── comorbidities_synthetic.csv     # Comorbidity records
├── medications_synthetic.csv       # Medication prescriptions
├── labs_synthetic.csv              # Laboratory test results
└── dataset_stats.json              # Summary statistics

---

Usage

Loading the Dataset

from datasets import load_dataset

# Load metadata with skin-tone annotations
dataset = load_dataset("kossiso-royce/africanized-skin-cancer", data_files="data/metadata.csv")

# Load synthetic EHR components
patients = load_dataset("kossiso-royce/africanized-skin-cancer", data_files="data/patients_synthetic.csv")
comorbidities = load_dataset("kossiso-royce/africanized-skin-cancer", data_files="data/comorbidities_synthetic.csv")
medications = load_dataset("kossiso-royce/africanized-skin-cancer", data_files="data/medications_synthetic.csv")
labs = load_dataset("kossiso-royce/africanized-skin-cancer", data_files="data/labs_synthetic.csv")

Example: Merge Patient Data with Metadata

import pandas as pd

metadata = pd.read_csv("data/metadata.csv")
patients = pd.read_csv("data/patients_synthetic.csv")

# Merge on lesion_id
full_data = metadata.merge(patients, on="lesion_id", how="left")

# Filter for dark skin patients with HIV
dark_skin_hiv = full_data[
    (full_data['phototype_bucket'] == 'Fitzpatrick V–VI (dark)')
].merge(
    pd.read_csv("data/comorbidities_synthetic.csv"),
    on="patient_id"
).query("condition == 'HIV'")

print(f"Dark skin patients with HIV: {len(dark_skin_hiv)}")

---

Known Limitations

1. Severe skin-tone bias: 88.9% light skin (Fitzpatrick I–II); only 4.3% dark skin
2. Provisional annotations: Automated skin-tone estimates require dermatologist validation
3. Synthetic EHR: Generated data, not real patient records; may not capture all clinical nuances
4. Geographic aggregation: African priors are averaged; regional variation not fully modeled
5. European diagnosis distribution: Reflects HAM10000's cohort; African melanoma patterns differ

---

Intended Use

✅ Appropriate Use

• Training/evaluating skin cancer detection models with fairness awareness
• Bias audits and representation gap analysis
• Multimodal learning research (image + EHR)
• Educational purposes and algorithm development

❌ Out-of-Scope

• Clinical deployment without validation on real African patient data
• Replacing real-world data collection efforts
• Fine-grained geographic/ethnic subgroup analysis

---

Ethical Considerations

• Transparency: All EHR fields flagged as source='synthetic'
• Bias documentation: Skin-tone imbalance explicitly reported
• Privacy: Original HAM10000 images de-identified; synthetic records contain no real PHI
• Fairness: Dataset highlights representation gaps; does not claim solved fairness

---

Citation

If you use this dataset, please cite:

@dataset{africanized_ham10000_2025,
  author = {Royce, Kossiso},
  title = {Africanized Skin Cancer Multimodal Dataset},
  year = {2025},
  publisher = {Hugging Face},
  howpublished = {\url{https://huggingface.co/datasets/kossiso-royce/africanized-skin-cancer}}
}

Original HAM10000:

@article{tschandl2018ham10000,
  title={The HAM10000 dataset, a large collection of multi-source dermatoscopic images of common pigmented skin lesions},
  author={Tschandl, Philipp and Rosendahl, Cliff and Kittler, Harald},
  journal={Scientific data},
  volume={5},
  number={1},
  pages={1--9},
  year={2018},
  publisher={Nature Publishing Group}
}

---

License

• Images: HAM10000 license (CC-BY-NC 4.0)
• Synthetic EHR: CC-BY 4.0 (attribution required)
• Metadata/Annotations: CC-BY 4.0

---

Contact

Maintainer: kossiso Royce
Issues: Please open an issue on the Hugging Face dataset page

---

Acknowledgments

• HAM10000 authors (Tschandl et al.) for the original dataset
• African dermatology literature for epidemiology priors
• Open-source tools: pandas, scikit-learn, OpenCV