README.md

metadata

tags:
  - synthetic
  - synthetic-data
  - tabular-classification
  - traditional-medicine
  - safety
  - africa
  - healthcare
  - pharmacovigilance
task_categories:
  - tabular-classification
  - other
task_ids:
  - tabular-multi-class-classification
language: []
pretty_name: Traditional Medicine Safety Dataset
size_categories:
  - 10K<n<100K

Traditional Medicine Safety Dataset

Description

A synthetic tabular dataset for traditional and complementary medicine safety in African populations. Models safety risks in the most widely used healthcare system on the continent.

Dataset Statistics

Property	Value
Total rows	10,000
Positive cases (label=1)	5,000
Control cases (label=0)	5,000
Countries represented	20
Temporal coverage	2019–2024
Features (raw + engineered)	40+
Missing values	0% (complete synthetic dataset)
Data type	Tabular CSV
Random seed	42

Class Balance & Distribution

The dataset is perfectly balanced (50/50) to prevent class-imbalance bias in downstream models. Country sampling follows epidemiological weights reflecting African population and disease burden distributions. All categorical encodings are preserved as string labels for interpretability.

Research Gap

80% of Africans use traditional medicine but there is virtually no structured safety data. Herb-drug interactions, hepatotoxicity, practitioner regulation, dosage standardisation, and paediatric vulnerability are all critically underdocumented.

African Healthcare Context

• 80% use traditional medicine as primary care
• <10% of healers are registered
• Hepatotoxicity is a leading cause of acute liver failure
• Limited integration policies outside South Africa, Ghana, Nigeria
• Bioprospecting concerns complicate research

Intelligence Sources

Source	URL
WHO TM Strategy	https://www.who.int/publications/i/item/9789240033895
African Union	https://au.int/
SA TMC	https://tmc.co.za/
NMEDA Nigeria	https://nmeda.gov.ng/

Columns

Column	Type	Description
country	string	Country
patient_age	int	Age
gender	string	Gender
practitioner_type	string	Type
years_practicing	int	Experience
reason_for_visit	string	Condition
herb_type	string	Herb type
preparation_method	string	Preparation
administration_route	string	Route
dosage_known	int	Known
frequency_daily	int	Frequency
duration_weeks	int	Duration
concurrent_western_meds	string	Concurrent
disclosure_to_doctor	int	Disclosure
adverse_event	string	Event
severity	string	Severity
time_to_onset_days	int	Onset
outcome	string	Outcome
hospitalisation_required	int	Hospital
previous_adverse_events	int	Prior
registration_status	int	Registered
quality_control	int	QC
label	int	1 = adverse, 0 = safe

Engineered Features

Feature	Description
practitioner_experience_score	Years + registration + QC
herb_risk_score	Prep + route + dosage + frequency
concurrent_med_risk	Interaction severity
care_integration_score	Disclosure - risk
patient_vulnerability	Age + prior events
event_severity_score	Weighted severity
high_risk_traditional	Composite flag

Feature Engineering Methodology

Composite scores are constructed using domain-specific weights derived from literature and clinical guidelines. Each score is rounded to 2 decimal places for reproducibility. Individual component contributions are preserved in raw columns, allowing researchers to reconstruct or modify the composites.

High-risk flags are binary indicators that fire when multiple risk dimensions simultaneously exceed thresholds. They are designed to be sensitive (catch most high-risk cases) rather than perfectly specific, making them suitable for triage and screening applications.

Feature Importance Notes

Based on preliminary Random Forest analysis:

• Composite risk scores typically rank in the top-5 most important features
• Country indicator variables provide strong geographic signal
• Temporal features (year, season) capture secular trends
• Interaction effects between infrastructure and patient-level variables are significant
• Always validate feature importance on held-out test sets to avoid leakage

Supported Use Cases

• Adverse event prediction
• Herb-drug interaction modelling
• Practitioner risk profiling
• QC benchmarking
• Integration policy design
• Paediatric safety
• Regulatory evaluation

Advanced Modelling Approaches

• Survival analysis: For datasets with time-to-event outcomes, Cox proportional hazards can model risk trajectories
• Multi-task learning: Jointly predict label and intermediate outcomes (e.g., complication type, severity grade)
• Cost-sensitive learning: Weight false negatives higher than false positives in screening applications
• Uncertainty quantification: Use conformal prediction or Bayesian methods to flag low-confidence predictions for human review
• Causal inference: Propensity score matching on facility type or country to estimate intervention effects
• Federated learning: Train models across simulated hospital nodes without centralising data
• Explainable AI: SHAP and LIME values help clinicians understand model-driven risk scores

Usage

from datasets import load_dataset

dataset = load_dataset("electricsheepafrica/africa-traditional-medicine-safety", split="train")
df = dataset.to_pandas()

import pandas as pd
from sklearn.model_selection import train_test_split
from sklearn.ensemble import RandomForestClassifier
from sklearn.metrics import classification_report, roc_auc_score

df = pd.read_csv("data/processed/traditional_features.csv")
X = df.select_dtypes(include=["int", "float"]).drop(columns=["label"])
y = df["label"]
X_train, X_test, y_train, y_test = train_test_split(X, y, test_size=0.2, stratify=y, random_state=42)
clf = RandomForestClassifier(random_state=42)
clf.fit(X_train, y_train)
print(classification_report(y_test, clf.predict(X_test)))
print("ROC-AUC:", roc_auc_score(y_test, clf.predict_proba(X_test)[:, 1]))

Data Generation

1. Positive cases with adverse events and risky use
2. Controls with safe use and known dosages
3. Leakage filtering for no events
4. Balanced 5,000 + 5,000
5. Experience, risk, and vulnerability features
6. Seed 42

Preprocessing Recommendations

1. One-hot encode categorical columns (country, facility type, region, etc.)
2. Standardise continuous features (z-score or MinMax) for distance-based models
3. Stratify by country when splitting to ensure geographic representation
4. Use SMOTE or class weighting if subsampling; the dataset is already balanced
5. Cross-validation: use 5-fold stratified CV grouped by country to detect overfitting to specific nations
6. Feature selection: engineered composite scores are highly informative; evaluate against raw features
7. Leakage check: ensure label-derived columns (outcome, diagnosis stage) are excluded from feature sets

Baseline Performance Expectations

Model	Expected Accuracy	Expected ROC-AUC	Notes
Logistic Regression	0.72–0.78	0.78–0.84	Good interpretability baseline
Random Forest	0.82–0.88	0.88–0.93	Handles non-linear interactions well
XGBoost / LightGBM	0.85–0.91	0.91–0.95	Best tabular performance
Neural Network (MLP)	0.80–0.86	0.85–0.90	Requires scaling; risk of overfitting
Linear SVM	0.74–0.80	0.80–0.85	Sensitive to scaling

These are approximate ranges on a stratified train/test split (80/20). Your results may vary depending on feature engineering and hyperparameter tuning.

Statistical Properties

• Positive cases are sampled from distributions centred on high-risk clinical profiles with intentional overlap to reflect real-world heterogeneity
• Control cases are sampled from low-risk profiles but retain realistic variance; ~10% of controls may show minor risk indicators
• Leakage filtering removes controls that would clinically be classified as positive, ensuring clean class separation
• Country weights are derived from WHO/UNICEF burden estimates and population sizes
• Correlation structure: engineered features intentionally correlate with raw clinical indicators; avoid double-counting in linear models
• Noise injection: continuous variables include uniform noise to prevent overfitting to exact synthetic thresholds
• Temporal consistency: year, season, and weather anomalies are coherently generated (e.g., drought months correlate with yield reductions)

Validation Checklist

Before using this dataset for research or production:

• Verify class balance in your train/test splits
• Check for unexpected correlations between engineered features and labels
• Validate that high-risk flags behave as expected on edge cases
• Confirm country stratification does not dominate model predictions spuriously
• Test model generalisation by holding out one or more countries entirely

Limitations

• Synthetic data
• Simplified herb categories
• Binary outcome

Ethical Considerations

• Respect traditional knowledge systems
• Avoid stigmatising healers or users
• Support integration
• Community consent
• Protect practitioner identities

Data Governance & Protection

• Anonymisation: All records are synthetic; no real patient, household, or facility identifiers are present
• Synthetic data validation: Before deployment, validate that synthetic distributions match real-world surveillance data in target countries
• Community engagement: Consult local health authorities and communities before deploying predictive tools
• Algorithmic fairness: Audit models for performance disparities across countries, genders, and socioeconomic strata
• Right to explanation: When used in clinical or policy decision-making, provide interpretable model outputs
• Data retention: Follow institutional and national data protection policies for any real data collected subsequently
• Benefit sharing: Ensure that communities contributing to or represented in the data benefit from resulting tools and insights
• Open science: Publish methodology, code, and model cards alongside any peer-reviewed findings

Recommended Splits

• Train: 70%
• Validation: 15%
• Test: 15%

Citation

@dataset{traditional_medicine_africa_2024,
  title = {Traditional Medicine Safety Dataset},
  author = {Electric Sheep Africa},
  year = {2024},
  url = {https://huggingface.co/datasets/electricsheepafrica/africa-traditional-medicine-safety}
}

License

CC BY-SA 4.0

Contact

electricsheepafrica@proton.me

Version History

• v1.0 — Initial release