Ssentongo/african-diabetes-dataset

African Diabetes Synthetic Dataset

Type 2 Diabetes & Gestational Diabetes Prediction

Version: 1.0
Release Date: November 2024
Context: Sub-Saharan Africa (24M adults with diabetes, projected 55M by 2045)
License: Research & Educational Use

---

Abstract

We present synthetic datasets for Type 2 Diabetes Mellitus (T2DM) and Gestational Diabetes Mellitus (GDM) modeling in Sub-Saharan Africa, addressing the rising NCD epidemic. With diabetes prevalence of 4-7% (urban: 8-12%) and 60-70% undiagnosed, early screening is critical. The datasets incorporate established risk factors with documented ORs: obesity BMI >30 (OR 7.19), family history (OR 3.50), previous GDM (OR 7.43), physical inactivity (OR 2.20), and hypertension (OR 2.50). For GDM, additional factors include previous macrosomia (OR 6.40), PCOS (OR 3.70), and maternal age >35 (OR 3.75). Seven datasets (25,000 samples total, 8.5 MB) provide configurations for population screening, urban high-risk cohorts, GDM antenatal prediction, and complication modeling. Models trained on these data are expected to achieve AUC-ROC >0.83 for diabetes risk prediction and >0.78 for complication onset, serving as proof-of-concept for primary care screening algorithms and antenatal GDM programs in African settings where diabetes remains severely underdiagnosed and undertreated.

Keywords: Type 2 Diabetes, Gestational Diabetes, GDM, Screening, Prevention, Complications, African Health, NCDs

---

Key Features

Risk Factors (Evidence-Based ORs)

Type 2 Diabetes:

• Obesity (BMI >30): OR 7.19 (strongest modifiable risk)
• Family history (1st degree): OR 3.50
• Previous GDM: OR 7.43
• Physical inactivity: OR 2.20
• Hypertension: OR 2.50
• Age >45 years: OR 2.75
• Urbanization: OR 2.40
• Overweight (BMI 25-30): OR 2.89

Gestational Diabetes:

• Previous GDM: OR 13.20 (strongest predictor)
• BMI >35: OR 8.56
• Previous macrosomia (baby >4kg): OR 6.40
• Family history diabetes: OR 4.90
• Maternal age >35: OR 3.75
• PCOS: OR 3.70
• HIV infection (African context): OR 1.65

Clinical Parameters

• Fasting glucose: Normal <100, Prediabetes 100-125, Diabetes ≥126 mg/dL
• HbA1c: Normal <5.7%, Prediabetes 5.7-6.4%, Diabetes ≥6.5%
• Lipid profiles: Total cholesterol, LDL, HDL, triglycerides
• Renal function: Creatinine, microalbuminuria (nephropathy marker)

Complications Modeled (5-10 Years Post-Diagnosis)

• Diabetic retinopathy: 25-35% at diagnosis → 60-80% by 20 years
• Diabetic nephropathy: 20-40% develop microalbuminuria
• Diabetic neuropathy: 30-50% prevalence
• Cardiovascular disease: OR 2-4 vs non-diabetics
• Diabetic foot ulcers: 15-25% lifetime risk

GDM Maternal/Fetal Outcomes

• Preeclampsia: OR 4.00 (synergy with our preeclampsia dataset!)
• Cesarean delivery: OR 1.50
• Macrosomia: OR 3.23
• Neonatal hypoglycemia: OR 7.40
• NICU admission: OR 2.30
• Progression to T2DM: 35-50% within 5 years

African Context

• Undiagnosed rate: 60-70% (vs 30% globally)
• Late presentation: 50-60% diagnosed with complications already present
• Medication access: Only 40-60% on any diabetes medication
• HbA1c monitoring: <30% have regular testing
• Good glycemic control (HbA1c <7%): Only 20-30% of patients
• Traditional medicine use: 40-60% alongside/instead of conventional treatment

---

Dataset Collection

Dataset Inventory

Dataset	N	Diabetic	Prediabetic	Diagnosed	Complications	Use Case
diabetes_t2dm_africa_baseline_1000	1,000	754 (75%)	126 (13%)	254 (34%)	46 (6%)	Prototyping
diabetes_t2dm_africa_large_5000	5,000	3,703 (74%)	685 (14%)	1,272 (34%)	200 (5%)	Main training
diabetes_t2dm_africa_extra_large_10000	10,000	7,458 (75%)	1,339 (13%)	2,559 (34%)	438 (6%)	Deep learning
diabetes_t2dm_africa_urban_high_risk_2000	2,000	1,494 (75%)	255 (13%)	529 (35%)	87 (6%)	Urban screening
diabetes_t2dm_africa_test_2000	2,000	1,477 (74%)	280 (14%)	522 (35%)	90 (6%)	Validation
diabetes_gdm_africa_antenatal_2000	2,000	1,232 (62%)	351 (18%)	429 (35%)	36 (3%)	GDM screening
diabetes_gdm_africa_large_5000	5,000	3,041 (61%)	899 (18%)	1,075 (35%)	106 (4%)	GDM training

Note: Prevalence enriched to enable ML training. T2DM baseline prevalence 7%, urban 13%, GDM 10%.

---

Expected Model Performance

Diabetes Risk Prediction

• Logistic Regression: AUC 0.79-0.83
• Random Forest: AUC 0.83-0.89
• XGBoost: AUC 0.86-0.92

Prediabetes Detection

• Multi-class F1: 0.72-0.80

Complication Prediction (Diagnosed Diabetics)

• Retinopathy/Nephropathy: AUC 0.76-0.84
• CVD risk: AUC 0.78-0.86

GDM Prediction

• Antenatal screening: AUC 0.81-0.88
• Adverse outcomes: AUC 0.74-0.82

Top Predictive Features

1. BMI/obesity status
2. Family history of diabetes
3. Age
4. Previous GDM (women)
5. Physical activity level
6. Hypertension
7. Urban vs rural residence
8. Fasting glucose (if available)
9. Previous macrosomia (GDM)
10. Parity (GDM)

---

Usage Example

import pandas as pd
from sklearn.ensemble import RandomForestClassifier

# Load data
train = pd.read_csv('diabetes_t2dm_africa_large_5000.csv')
test = pd.read_csv('diabetes_t2dm_africa_test_2000.csv')

# Diabetes Risk Prediction
features = ['age', 'bmi', 'family_history_diabetes', 
            'physically_active', 'has_hypertension',
            'residence', 'previous_gdm']

# Encode categoricals
from sklearn.preprocessing import LabelEncoder
for col in ['residence']:
    le = LabelEncoder()
    train[col] = le.fit_transform(train[col])
    test[col] = le.transform(test[col])

# Binary target: Diabetic vs Non-diabetic
X_train = train[features]
y_train = (train['diabetes_status'] == 'Diabetic').astype(int)

# Train
model = RandomForestClassifier(n_estimators=100, class_weight='balanced')
model.fit(X_train, y_train)

# Expected AUC: 0.83-0.89

---

Clinical Applications

Primary Care Screening

• Risk stratification: Identify high-risk individuals for OGTT
• Resource allocation: Target limited glucose testing capacity
• Prevention programs: Lifestyle intervention prioritization

Antenatal Care (GDM)

• Early screening: Identify women needing OGTT at booking
• Intervention timing: Optimize diet/medication initiation
• Postpartum follow-up: T2DM progression prediction

Complication Prevention

• Retinopathy screening: Identify patients needing ophthalmology referral
• Nephropathy monitoring: Prioritize microalbuminuria testing
• Foot care programs: Target high-risk patients

Public Health Planning

• NCD burden estimation: Forecast diabetes prevalence by region
• Medication supply: Metformin, insulin demand forecasting
• HbA1c lab capacity: Testing scale-up planning

---

Limitations

What This Dataset IS:

• ✅ Population screening risk stratification
• ✅ Prevention program targeting (lifestyle, metformin)
• ✅ GDM antenatal prediction modeling
• ✅ Complication risk assessment

What This Dataset IS NOT:

• ❌ Replacement for OGTT (gold standard diagnosis)
• ❌ Including continuous glucose monitoring data
• ❌ Capturing medication side effects
• ❌ Modeling dietary/cultural factors precisely

Mandatory Next Steps:

1. Prospective validation in African primary care
2. Integration with existing NCD screening programs
3. Calibration to region-specific prevalence
4. Ethical review for clinical algorithm deployment

---

Citation

African Diabetes Synthetic Dataset (2024)
Type 2 Diabetes and Gestational Diabetes prediction modeling
Version 1.0, Generated November 2024
Organization: Electric Sheep Africa
License: CC-BY-4.0

Primary Sources:

• IDF Diabetes Atlas 2021 (Africa regional data)
• Atun et al., Lancet Diabetes Endocrinol 2017 (SSA diabetes)
• WHO GDM Diagnostic Criteria 2013
• Macaulay et al., Int J Gynaecol Obstet 2014 (GDM Africa)
• WINGS Study (Nigeria, Ghana, Kenya cohorts)

---

Quick Reference

T2DM Risk Features (7):

features = ['age', 'bmi', 'family_history_diabetes',
            'physically_active', 'has_hypertension',
            'previous_gdm', 'residence']

GDM Risk Features (9):

features = ['age', 'bmi', 'family_history_diabetes',
            'previous_gdm', 'parity', 'previous_macrosomia',
            'pcos', 'hiv_positive', 'residence']

Targets:

• Classification: diabetes_status (Normal/Prediabetes/Diabetic)
• Glucose values: fasting_glucose_mg_dl, hba1c_percent
• Complications: retinopathy, nephropathy, neuropathy, cardiovascular_disease

Expected Performance:

• Diabetes risk: AUC 0.83-0.92
• GDM prediction: AUC 0.81-0.88
• Complications: AUC 0.76-0.86

---

Status: Research Use Only - Prospective Validation Required
Last Updated: November 7, 2024
Contact: https://huggingface.co/electricsheepafrica