electricsheepafrica/african-malaria-dataset

African Malaria Synthetic Dataset

A Literature-Informed Probabilistic Approach to Malaria Detection and Severity Prediction

Version: 1.0
Release Date: November 2024
Context: Sub-Saharan African Epidemiology
License: Research & Educational Use

---

Abstract

We present synthetic datasets for malaria detection and severity prediction in Sub-Saharan Africa, generated using literature-informed probabilistic modeling based on 2024 epidemiological research. Malaria remains the leading cause of childhood mortality in Africa, with 263 million cases globally in 2023, predominantly in Sub-Saharan countries. The datasets incorporate age-specific distributions (Log-normal for children, Weibull for adults), parasitemia levels, clinical presentations, and risk factors documented in recent peer-reviewed studies from Ethiopia, Tanzania, and multi-country systematic reviews. With estimated incidence rates of 60.4 per 1,000 population at risk and mortality rates of 9.33 per 1,000 children annually, early detection and severity assessment are critical for treatment decisions. Our synthetic generation approach enables algorithm development for resource-limited settings where longitudinal data collection faces substantial barriers. Seven datasets (varied sizes and endemic scenarios) provide configurations for diagnostic algorithms, severity prediction, and seasonal variation analysis. These data support development of ML models for malaria screening, parasitemia estimation, and severe case identification—serving as proof-of-concept for deployment in African healthcare settings.

Keywords: Malaria, P. falciparum, Synthetic Data, African Health, Machine Learning, Parasitemia, Severe Malaria, Low-Resource Settings

---

1. Introduction

1.1 Clinical Context

Malaria, caused primarily by Plasmodium falciparum in Africa, affects over 263 million people globally, with the African region bearing 94% of cases and 95% of deaths (WHO 2024). Key epidemiological features include:

• Age vulnerability: Children under 5 account for 80% of malaria deaths
• Endemic transmission: Incidence rates reach 400+ per 100,000 in peak countries
• Preventable mortality: 9.33 deaths per 1,000 children annually in stable endemic areas
• Seasonal variation: 60% of cases occur during rainy seasons

Early detection and accurate severity assessment are critical for appropriate treatment (artemisinin-based therapy vs. parenteral artesunate for severe cases).

1.2 Data Collection Challenges

Real-world malaria dataset construction in African settings faces:

• Diagnostic limitations: Microscopy requires skilled technicians; RDTs have sensitivity issues
• Parasitemia quantification: Labor-intensive microscopy; automated counters rare
• Seasonal variation: Data collection must span multiple transmission seasons
• Healthcare access: Rural populations (85.7% of cases) have delayed presentations
• Resource constraints: Limited laboratory capacity for culture confirmation (31.8% rate)

1.3 Synthetic Data Rationale

We employ literature-informed synthetic generation as a scaffold for:

1. Diagnostic algorithm development: Severity prediction models without waiting for longitudinal cohorts
2. Resource allocation: Demonstrate ML feasibility for funding applications
3. Clinical decision support: Identify critical features (parasitemia, hemoglobin, age) for risk stratification
4. Training optimization: Balance class distributions for rare severe cases (5-15%)

This approach accelerates deployment-ready tools while real validation studies are conducted.

---

2. Methodology

2.1 Generation Framework

Probabilistic Sampling with Epidemiological Constraints

We extract statistical distributions from published meta-analyses and cohort studies, implementing Monte Carlo sampling:

For each sample i:
  1. Age_i ~ Bimodal(Children: LogNormal(μ=1.91, σ=1.20), 
                     Adults: Weibull(k=1.01, λ=11.89))
  2. Sex_i ~ Bernoulli(p_male = 0.589)
  3. Risk_factors_i ~ {Residence, Season, Mosquito_net}
  4. P(Malaria|features_i) = f(Age, Residence, Season, Protection)
  5. Malaria_i ~ Bernoulli(P(Malaria|features_i))
  6. If Malaria_i:
       - Parasitemia_level ~ Categorical([0.238, 0.631, 0.131])
       - Species ~ Categorical([0.785, 0.161, 0.054])
       - Hemoglobin ~ Normal(age_specific_mean - parasitemia_effect)
       - Severe ~ f(Age, Parasitemia, Species)

2.2 Age Distribution Parameters

Derived from Global Burden of Disease 2019 + Malaria Journal 2023

Age follows bimodal distribution reflecting African epidemiology:

Age Group	% Distribution	Statistical Model	Parameters
0-2 years	15.4%	Log-normal	μ = 1.9144, σ = 1.1995
2-6 years	30.5%	Log-normal	(same parameters, age-capped)
6-12 years	17.6%	Weibull	k = 1.007, λ = 11.8898
≥12 years	36.5%	Weibull	(shifted distribution)

Rationale: Young children lack immunity; all ages susceptible in endemic areas.

2.3 Malaria Probability Model

Base prevalence adjusted by epidemiological risk factors:

P_base = malaria_prevalence  # 0.20 (low) to 0.60 (hyperendemic)

# Age effects (immunological vulnerability)
if age < 5: P_base *= 1.5
elif age < 15: P_base *= 1.2

# Geographic risk (transmission intensity)
if residence == 'Rural': P_base *= 1.3
else: P_base *= 0.7  # Urban: lower transmission

# Seasonal variation
if season == 'Rainy': P_base *= 1.4
else: P_base *= 0.7

# Preventive measures
if uses_mosquito_net: P_base *= 0.5  # 50% efficacy

P_final = min(P_base, 0.95)

2.4 Clinical Parameters

Parasitemia Levels (Ethiopian Study 2022)

Distribution: Low 23.8%, Moderate 63.1%, High 13.1%

Level	Parasites/μL Range	Clinical Significance
Low	50 - 5,000	Asymptomatic/mild symptoms
Moderate	5,001 - 100,000	Uncomplicated malaria
High	100,001 - 500,000	Risk of severe complications

Hyperparasitemia (>200,000/μL): 2-3× mortality risk

Plasmodium Species (Africa-specific)

• P. falciparum: 78.5% (cerebral malaria risk, drug resistance)
• P. vivax: 16.1% (relapsing, East Africa)
• Mixed infection: 5.4% (complicated treatment)

Hemoglobin Correlation

Malaria-induced anemia:

• Low parasitemia: Hb drop 1.5 ± 0.5 g/dL
• Moderate: Hb drop 2.5 ± 0.8 g/dL
• High: Hb drop 4.0 ± 1.2 g/dL

Severe anemia (Hb <7 g/dL): Transfusion required, 30% of severe cases

2.5 Severe Malaria Criteria

WHO Definition (one or more):

• Cerebral malaria (20% of severe cases in our model)
• Respiratory distress (15%)
• Shock/circulatory collapse (10%)
• Acute kidney injury (12%, adults predominantly)
• Severe anemia (Hb <7 g/dL)
• Hyperparasitemia (>10% infected RBCs or >200,000/μL)

Incidence: 5% baseline, increased to 15-40% with risk factors:

• Age <5 years: 3× risk
• High parasitemia: 2.5× risk
• Parasitemia >200,000/μL: Additional 1.5× risk

2.6 Mortality Rates

Age-stratified (from NCBI Disease and Mortality SSA):

Age Group	Rate per 1,000/year	% of Age Group Deaths
<5 years	9.33	28.2%
5-14 years	1.58	52.2%
≥15 years	0.60	6.0%

Modifiers:

• Severe malaria: 10× baseline risk
• Hyperparasitemia: 2× additional risk
• Delayed treatment: >3 days fever → 1.5× risk

2.7 Feature Set

28 features across six categories:

• Demographics (6): patient_id, age_years, age_months, age_group, sex, residence
• Epidemiological (2): season, uses_mosquito_net
• Laboratory (5): malaria_status, parasitemia_level, parasitemia_count, plasmodium_species, hemoglobin_g_dl
• Anemia (1): anemia_status (None/Moderate/Severe)
• Clinical Symptoms (7): fever_days, has_fever, has_chills, has_headache, has_vomiting, has_diarrhea, has_weakness
• Severe Complications (5): severe_malaria, cerebral_malaria, respiratory_distress, shock, acute_kidney_injury
• Outcome (2): outcome (Healthy/Treated/Hospitalized/Died), malaria_probability_score

Target variables:

• Classification: malaria_status (Positive/Negative)
• Severity: severe_malaria (Boolean)
• Outcome: outcome (4-class)

---

3. Dataset Collection

3.1 Dataset Inventory

Seven datasets provide varied experimental configurations:

Dataset	N	Malaria+	%	Endemic Level	Use Case
malaria_ssa_baseline_1000	1,000	~400	40%	High endemic	Rapid prototyping
malaria_ssa_large_5000	5,000	~2,000	40%	High endemic	Main training
malaria_ssa_extra_large_10000	10,000	~4,000	40%	High endemic	Deep learning
malaria_ssa_low_endemic_2000	2,000	~400	20%	Low endemic	Low transmission areas
malaria_ssa_hyperendemic_2000	2,000	~1,200	60%	Hyperendemic	Peak transmission zones
malaria_ssa_seasonal_rainy_1000	1,000	~400	40%	Rainy season	Seasonal modeling
malaria_ssa_test_2000	2,000	~400	40%	Hold-out validation	Independent test set

Critical: Test dataset uses different random state and must never be used for training.

3.2 Class Distribution Analysis

Typical high-endemic dataset (baseline_1000):

• Negative cases: ~600 (60%)
• Positive cases: ~400 (40%)
  • Uncomplicated: ~340 (85%)
  • Severe: ~60 (15%)
  • Deaths: ~5-10 (1-2% of positive)

Imbalance considerations:

• Severe malaria is minority class (15% of positives)
• Deaths are extremely rare (requiring oversampling or focal loss)
• Age stratification: Under-5 overrepresented in severe cases

3.3 Quality Control

Validation checks implemented in generator:

1. Age-parasitemia correlation: Children <5 have 30% higher mean parasitemia
2. Hemoglobin-parasitemia inverse relationship: Pearson r ≈ -0.65
3. Species-geography consistency: P. falciparum dominance (78.5%)
4. Mortality-severity alignment: No deaths without elevated risk factors
5. Seasonal variation: Rainy season 1.4× prevalence vs. dry

3.4 Feature Importance (Expected)

Based on epidemiological literature:

Top predictors for malaria diagnosis:

1. Fever presence/duration (AUC ~0.85 alone)
2. Age group (children <5 vs. adults)
3. Season + mosquito net (preventive behaviors)
4. Residence (rural vs. urban)

Top predictors for severity:

1. Parasitemia count (most critical)
2. Age <5 years
3. Hemoglobin level
4. Days to presentation

---

4. Experimental Considerations

4.1 Recommended Train/Test Splits

Option A: Single endemic level

• Train: malaria_ssa_large_5000 (5,000 samples)
• Validation: 20% holdout from training set
• Test: malaria_ssa_test_2000 (2,000 samples, independent seed)

Option B: Multi-endemic training

• Train: Combine high_endemic + low_endemic (7,000 samples)
• Test: hyperendemic_2000 (transfer learning evaluation)

Option C: Temporal simulation

• Train: seasonal_rainy_1000 (dry season held out)
• Test: Synthetic "dry season" variant (adjusted prevalence)

4.2 Model Architecture Suggestions

Task 1: Malaria Detection (Binary Classification)

Baseline: Logistic Regression with age, fever, season, residence
Target: AUC-ROC > 0.85

Advanced: XGBoost/Random Forest with full feature set
Target: AUC-ROC > 0.92, Sensitivity > 0.95 (critical for screening)

Task 2: Severity Prediction (Binary on Positive Cases)

Baseline: Decision Tree on parasitemia + age + hemoglobin
Target: AUC-ROC > 0.80

Advanced: Gradient Boosting with interaction terms
Target: AUC-ROC > 0.90, Specificity > 0.85 (avoid unnecessary referrals)

Task 3: Outcome Prediction (4-class)

Advanced: Multi-class XGBoost or Neural Network
Evaluation: Macro F1-score, weighted by class importance

4.3 Class Imbalance Mitigation

Severe malaria (15% of positive cases):

• SMOTE: Synthetic minority oversampling
• Focal Loss: γ=2 to down-weight easy negatives
• Class weights: Inverse frequency weighting (1:6 severe:uncomplicated)

Deaths (<2% of samples):

• Stratified k-fold: Preserve death cases in each fold
• Ensemble methods: Boosting to focus on hard cases
• Threshold tuning: Optimize F1 at cost-sensitive threshold

4.4 Feature Engineering Recommendations

Derived features to create:

1. Age-parasitemia interaction: age_under5 * log(parasitemia + 1)
2. Severity risk score: (parasitemia/100000) * (1 if age<5 else 0.5) * (15-hemoglobin)/5
3. Endemic index: residence_rural * season_rainy * (1-mosquito_net)
4. Anemia severity: Categorical encoding of Hb levels

Domain knowledge rules:

• No fever but high parasitemia → Asymptomatic (rare, model should learn)
• Fever >7 days → Likely treatment failure or co-infection
• Cerebral malaria without high parasitemia → Suspicious (data quality flag)

---

5. Validation & Limitations

5.1 Expected Model Performance

Realistic targets based on synthetic data quality:

Task	Metric	Expected Range	Clinical Threshold
Malaria detection	AUC-ROC	0.88-0.94	>0.90 for deployment
	Sensitivity	0.92-0.98	>0.95 (screening priority)
	Specificity	0.75-0.88	>0.80 (RDT parity)
Severity prediction	AUC-ROC	0.85-0.92	>0.85 for triage
	PPV	0.60-0.75	Balance false alarms

Note: Performance on real data expected to be 5-10% lower due to:

• Measurement noise (microscopy variability)
• Missing values (lab equipment failures)
• Co-infections (not modeled)
• Drug-resistant parasites (emerging, not in model)

5.2 Synthetic Data Limitations

Known simplifications:

1. Independence assumption: Risk factors generated independently (reality: clustered in families/villages)
2. No drug resistance: Assumes treatment-sensitive parasites
3. Single infection: No co-morbidities (HIV, malnutrition) modeled
4. Perfect measurement: No lab error, RDT false positives/negatives
5. Stable transmission: No epidemic spikes or intervention campaigns

Not suitable for:

• Policy cost-effectiveness modeling (no treatment pathways)
• Drug resistance surveillance (no genotype data)
• Entomological correlation (no vector data)

5.3 Real-World Validation Requirements

Before clinical deployment, models must be validated on:

1. Prospective cohort: ≥500 patients with confirmed microscopy + RDT
2. Multi-site: Rural + urban facilities (3+ sites)
3. Seasonal coverage: Both rainy and dry seasons
4. External validation: Different geographic region (test generalization)
5. Clinical impact: Diagnostic accuracy, time-to-treatment, cost per case

Ethical imperative: Synthetic models are research tools only until real-world safety/efficacy demonstrated.

---

6. Use Cases & Applications

6.1 Clinical Decision Support

Triage at point-of-care:

• Input: Age, fever duration, season, residence, recent travel
• Output: Malaria risk score (0-100%)
• Action: If >70% → RDT recommended, If >90% → Presumptive treatment

Severity prediction:

• Input: RDT+, parasitemia (if available), hemoglobin, age
• Output: Severe malaria probability
• Action: If >30% → Refer to hospital, If >50% → Immediate IV artesunate

6.2 Resource Allocation

Laboratory prioritization:

• Predict which RDT+ patients need microscopy (parasitemia quantification)
• Allocate limited reagents to high-risk cases

Seasonal preparedness:

• Train on rainy season data → Predict bed needs, drug stockpiles
• Early warning systems (integrate with weather data)

6.3 Research & Training

Algorithm comparison:

• Benchmark new ML methods on standardized datasets
• Reproducible experiments (fixed random seed)

Medical education:

• Teach malaria epidemiology with interactive data exploration
• Simulate outbreak scenarios for public health training

---

7. Implementation Guide

7.1 Quick Start

Generate default dataset:

cd Malaria/
python malaria_data_generator.py -n 1000 -p 0.40 -s 42
# Output: malaria_synthetic_ssa_YYYYMMDD_HHMMSS.csv

Generate all standard datasets:

python generate_datasets.py
# Generates 7 datasets with varied configurations

Parameters:

• -n, --num-samples: Dataset size (default: 1000)
• -p, --prevalence: Malaria prevalence 0.0-1.0 (default: 0.40)
• -o, --output: Custom filename (optional)
• -s, --seed: Random seed for reproducibility (optional)

7.2 Loading Data

Python (pandas):

import pandas as pd

# Load dataset
df = pd.read_csv('malaria_ssa_baseline_1000.csv')

# Binary classification task
X = df.drop(['patient_id', 'malaria_status', 'outcome'], axis=1)
y = (df['malaria_status'] == 'Positive').astype(int)

# Severity prediction (positive cases only)
df_positive = df[df['malaria_status'] == 'Positive']
X_severity = df_positive[['parasitemia_count', 'age_years', 'hemoglobin_g_dl']]
y_severity = df_positive['severe_malaria'].astype(int)

R:

library(tidyverse)

df <- read_csv("malaria_ssa_baseline_1000.csv")

# Preprocessing
df <- df %>%
  mutate(
    malaria = as.factor(if_else(malaria_status == "Positive", 1, 0)),
    age_group = as.factor(age_group),
    severe = as.factor(severe_malaria)
  )

7.3 Baseline Models

Scikit-learn pipeline:

from sklearn.ensemble import RandomForestClassifier
from sklearn.model_selection import cross_val_score
from sklearn.preprocessing import LabelEncoder

# Encode categorical variables
le = LabelEncoder()
X_encoded = X.copy()
for col in ['age_group', 'sex', 'residence', 'season']:
    X_encoded[col] = le.fit_transform(X[col])

# Train Random Forest
rf = RandomForestClassifier(n_estimators=100, max_depth=10, random_state=42)
scores = cross_val_score(rf, X_encoded, y, cv=5, scoring='roc_auc')
print(f"Mean AUC-ROC: {scores.mean():.3f} (+/- {scores.std():.3f})")

---

8. Citation & References

8.1 Citing This Dataset

@misc{malaria_synthetic_africa_2024,
  title={African Malaria Synthetic Dataset: Literature-Informed Probabilistic Modeling},
  author={[Your Team]},
  year={2024},
  note={Based on GBD 2019, WHO 2024, and African epidemiological studies}
}

8.2 Source Literature

Age distributions:

1. Malaria Journal (2023). "Estimated distribution of malaria cases among children in SSA by age." DOI: 10.1186/s12936-023-04811-z

Clinical parameters: 2. PMC9391188 (2022). "Malaria Infection, Parasitemia, and Hemoglobin Levels in Febrile Patients, Ethiopia" 3. NCBI NBK2286. "Disease and Mortality in Sub-Saharan Africa: Malaria Chapter"

Epidemiology: 4. WHO World Malaria Report (2024). Global prevalence, incidence, mortality data 5. GBD 2019 (IHME). Age-stratified burden estimates for Sub-Saharan Africa

Statistical distributions: 6. Log-normal parameters (μ=1.9144, σ=1.1995) from GBD meta-analysis 7. Weibull parameters (k=1.007, λ=11.8898) fitted to age-prevalence curves

---

9. Contact & Support

Dataset Issues: Report bugs or inconsistencies via project repository
Clinical Questions: Consult WHO malaria treatment guidelines
Model Development: Share results to improve future dataset versions

Disclaimer: This is synthetic data for research/training purposes only. NOT validated for clinical use. Real-world deployment requires prospective validation with ethical approval.

---

Last Updated: November 2024
Version: 1.0
Status: Initial release for algorithm development