Synthetic data generation and validation methodology
Epidemiological foundation and parameter justification
The synthetic dataset was systematically generated to emulate realistic malaria transmission patterns in rural Sub-Saharan African communities, based on established epidemiological literature and WHO surveillance data. All parameters were derived from peer-reviewed clinical studies and validated against published epidemiological benchmarks to ensure clinical relevance and methodological rigor.
Sample size justification The synthetic population size of \(N=10,100\) was selected to represent a realistic healthcare catchment area for rural health centers in the Gambella region, Ethiopia. This choice reflects the demographic and healthcare infrastructure reality of rural Sub-Saharan Africa, where health centers typically serve 15,000–25,000 people through networks of 5–7 health posts, each covering 3,000–5,000 individuals8,9.
Gambella region, with a total population of approximately 525,000 (75% rural)10,11, provides the demographic foundation for this sample size choice. The 10,100 individuals represent approximately 3.6% of a typical district population or the aggregated catchment reflecting documented healthcare service areas in rural Ethiopia12,13. This population size aligns with healthcare catchment areas where the majority of health centers serve 2,000–5,000 people within walking distance8.
From a statistical perspective, the sample size meets established guidelines for machine learning diagnostic studies (\(\ge 10,000\) samples for robust performance assessment) while providing sufficient positive cases (\(n=1,969\), 19.5% prevalence) and negative cases (\(n=8,131\)) for accurate diagnostic performance evaluation14,15. The sample represents the realistic patient volume that would utilize automated diagnostic tools in rural health centers, supporting both scientific validity and practical implementation relevance.
Population characteristics The synthetic population (\(N=10,100\)) was designed to represent a high-transmission malaria-endemic community with 20% prevalence, consistent with WHO estimates for rural Sub-Saharan Africa16. This prevalence level reflects areas with perennial transmission and limited vector control interventions, typical of resource-constrained settings where automated diagnostic tools would have greatest impact.
Demographic parameters Age distribution followed UN Population Division data for Sub-Saharan Africa17, with mean age 22.8 years (\(\text {SD}=16.5\)), reflecting the young population structure characteristic of these regions. Age ranges were constrained to 0.5–85 years to represent realistic clinical populations seeking malaria diagnosis. Demographic patterns were validated against Tanzania Demographic and Health Survey 2015-1618 to ensure regional representativeness.
Clinical parameter derivation and literature validation
Clinical manifestation probabilities were systematically derived from large-scale clinical studies and systematic reviews in malaria-endemic regions. Fever prevalence was established at 85% in malaria-positive cases, based on the comprehensive study by Roucher et al.19 involving 2,847 microscopy-confirmed cases in Senegal. Chills occurrence was documented at 78% prevalence, derived from the clinical cohort study conducted by Kahigwa et al.20 in Tanzania, while fatigue manifestation was observed in 82% of cases, based on the community survey by Mwangi et al.21 in coastal Kenya involving 1,844 participants.
For malaria-negative individuals, background symptom rates reflected the prevalence of other febrile illnesses commonly encountered in Sub-Saharan Africa. These baseline rates were established as fever occurring in 25% of cases, chills in 15%, and fatigue in 35%, based on comprehensive community health surveys conducted by Snow et al.22. These differential symptom probabilities ensure realistic representation of the diagnostic challenge in endemic areas where multiple febrile illnesses present with overlapping clinical features.
Symptom probabilities were carefully adjusted to incorporate age-specific immune responses consistently observed in endemic populations2,6. Children under five years demonstrated 15% higher symptom probability, reflecting their limited acquired immunity and increased vulnerability to severe manifestations. Elderly individuals over 65 years showed 10% higher symptom probability, representing the waning immunity that occurs with advanced age. Adults between 5 and 65 years exhibited 5% lower symptom probability, reflecting the partial acquired immunity that develops from repeated exposure to malaria parasites in endemic settings.
Environmental factor modeling and climate integration
Temperature Atmospheric temperature (\(^\circ\)C) was modeled using normal distribution (\(\mu =26.5\), \(\sigma =4.2\)) based on meteorological data from malaria-endemic regions6. A positive correlation with malaria status (+0.8\(^\circ\)C mean shift for positive cases) reflected optimal transmission temperatures (20–30\(^\circ\)C range) established by Paaijmans et al.23.
Rainfall Monthly rainfall (mm) followed gamma distribution (shape=1.8, scale=12.5) reflecting typical seasonal patterns in Sub-Saharan Africa23,24. This distribution captures the right-skewed nature of precipitation data and correlation with increased vector breeding sites and malaria transmission intensity.
Feature interaction modeling and biological realism
Unlike simple random generation approaches, the synthetic data incorporated clinically relevant interactions that reflect the complex biological and epidemiological relationships observed in real-world malaria transmission. Age-symptom interactions were modeled to capture the immune response maturation that affects symptom expression patterns across different age groups, ensuring that younger and older populations exhibited appropriate vulnerability profiles. Environmental-disease interactions incorporated the established relationships between temperature and rainfall patterns and malaria transmission probability, reflecting the well-documented influence of climatic factors on vector breeding and parasite development.
Symptom clustering patterns were implemented to ensure realistic co-occurrence of fever, chills, and fatigue, avoiding the unrealistic independence that would result from purely random generation. Additionally, seasonal effects were incorporated to model how environmental factors influence both symptom severity and disease progression, capturing the temporal dynamics that characterize malaria epidemiology in endemic regions. These interaction patterns ensure that the synthetic data maintains biological plausibility while providing a controlled framework for systematic algorithm comparison.
Data generation algorithm
The Monte Carlo simulation approach employed the following systematic process:
For each individual i in population \(N=10,100:\)
$$\begin{aligned} M_i&\sim \text {Bernoulli}(0.20) \end{aligned}$$
(1)
$$\begin{aligned} \text {Age}_i&\sim \mathcal {N}(22.8, 16.5^2) \text { [constrained to 0.5–85 years]} \end{aligned}$$
(2)
$$\begin{aligned} \text {Temp}_i&\sim \mathcal {N}(26.5 + 0.8 \times M_i, 4.2^2) \end{aligned}$$
(3)
$$\begin{aligned} \text {Rain}_i&\sim \text {Gamma}(1.8, 12.5) \end{aligned}$$
(4)
$$\begin{aligned} \text {AF}_i&= f(\text {Age}_i) \end{aligned}$$
(5)
$$\begin{aligned} \text {EF}_i&= f(\text {Temp}_i, \text {Rain}_i) \end{aligned}$$
(6)
$$\begin{aligned} P(\text {Fever}_i)&= \text {Base probability} \times \text {AF}_i \times \text {EF}_i \end{aligned}$$
(7)
$$\begin{aligned} P(\text {Chills}_i)&= \text {Base probability} \times \text {AF}_i \end{aligned}$$
(8)
$$\begin{aligned} P(\text {Fatigue}_i)&= \text {Base probability} \times \text {AF}_i \end{aligned}$$
(9)
All probabilities were constrained to biologically plausible range [0.05, 0.98].
Model selection and algorithmic diversity
Five machine learning models representing diverse algorithmic approaches were selected for comprehensive comparison to ensure robust evaluation across different computational paradigms. Naive Bayes was included as a probabilistic classifier that assumes feature independence, providing an interpretable baseline performance measure that reflects fundamental probabilistic relationships in the data6. Logistic Regression served as the linear model for estimating malaria probability, valued particularly for its clinical interpretability and widespread acceptance in medical applications3.
Random Forest was selected as an ensemble method capable of handling non-linear relationships and complex feature interactions, representing the tree-based ensemble approach to classification problems1. XGBoost was included as a gradient-boosted decision tree algorithm optimized for both performance and computational efficiency, representing the current state-of-the-art in gradient boosting methodology25. Finally, Enhanced Bayesian Logistic Regression was developed as a novel Bayesian approach incorporating clinical domain knowledge integration and uncertainty quantification, representing an advanced probabilistic framework that addresses both prediction accuracy and clinical interpretability requirements7,19.
Computational reproducibility framework
To ensure complete reproducibility and transparency, we implemented a comprehensive computational framework following open science best practices. All analyses were conducted in R version 4.4.3 with specific package versions documented in the repository requirements file. The analysis pipeline is organized as a streamlined 2-step process: (1) enhanced synthetic data generation with clinical validation protocols, and (2) integrated machine learning analysis including model development, statistical significance testing, and comprehensive visualization generation.
Statistical validation protocol Model performance comparisons employed McNemar’s test for pairwise accuracy differences, bootstrap resampling (\(n=1000\)) for confidence interval estimation, and Friedman test for overall model ranking significance. Cross-validation stability was assessed through coefficient of variation analysis across fold-specific performance metrics, demonstrating robust performance estimates with \(<3\)% variability.
Computational requirements Complete execution requires approximately 2.4 hours on standard academic computing infrastructure (\(\ge 8\)GB RAM, \(\ge 4\) CPU cores), with individual model training times ranging from 0.11 minutes (Logistic Regression) to 144.8 minutes (Enhanced Bayesian LR with MCMC sampling). The repository includes sample output files to enable immediate reviewer validation without requiring full analysis execution. All statistical tests, hyperparameters, and methodological specifications are fully documented and version-controlled through the public repository.
Experimental design and statistical framework
Data partitioning and cross-validation
Data splitting Stratified random sampling divided the dataset maintaining class distribution:
-
Training set: 70% (\(n=7,070\)) for model development and hyperparameter optimization
-
Validation set: 15% (\(n=1,515\)) for model selection and threshold tuning
-
Test set: 15% (\(n=1,515\)) for final performance evaluation
Cross-validation protocol Five-fold cross-validation on training data ensured robust parameter estimation and performance assessment, with stratification maintaining malaria prevalence across folds.
Computational efficiency The analysis pipeline is optimized for diverse computational environments, with individual model training times varying significantly: Logistic Regression (0.11 minutes), Naive Bayes (0.43 minutes), XGBoost (0.61 minutes), Random Forest (0.68 minutes), and Enhanced Bayesian LR (144.8 minutes). The Enhanced Bayesian analysis requires substantial computational time due to MCMC sampling for clinical domain knowledge integration, while remaining models provide rapid validation capabilities. For efficient peer review, the repository includes pre-generated sample outputs alongside full reproduction scripts, enabling both immediate validation and comprehensive methodological verification.
Cost-sensitive threshold optimization
Clinical cost framework Medical contexts prioritize sensitivity (avoiding missed diagnoses) over specificity. Cost weights were assigned based on clinical consequences:
-
False Negative cost (\(C_{FN}\)): 15 units (missed malaria diagnosis)
-
False Positive cost (\(C_{FP}\)): 3 units (unnecessary treatment)
-
Cost ratio: 5:1 reflecting clinical priority for sensitivity
The total cost function was defined as:
$$\begin{aligned} \text {Total Cost} = C_{FN} \times \text {FN} + C_{FP} \times \text {FP} \end{aligned}$$
(10)
Threshold selection Optimal classification thresholds were determined by minimizing total cost across probability thresholds [0.05, 0.95] with 0.05 increments, enabling clinical priority alignment.
Performance evaluation metrics
Primary metrics
-
Area under roc curve (AUC) Overall discriminative performance
-
Sensitivity Proportion of malaria cases correctly identified (clinical priority)
-
Specificity Proportion of non-malaria cases correctly identified
-
Total cost Weighted sum of false negatives and false positives
Secondary metrics
-
Accuracy Overall correct classification rate
-
Precision Positive predictive value
-
F1-score Harmonic mean of precision and sensitivity
-
Area under precision-recall curve (AUPRC) Performance under class imbalance
Statistical rigor and significance testing
Bootstrap confidence intervals 1000 bootstrap resamples provided robust confidence intervals for all performance metrics with empirical coverage assessment.
Multiple comparison framework
-
Friedman test Overall significance of model ranking differences
-
McNemar’s test Pairwise model comparisons for classification accuracy
-
Effect size analysis Cohen’s d for practical significance assessment
Cross-validation stability Coefficient of variation and reliability analysis across folds to assess performance consistency.
