Comparative analysis of machine learning models for malaria detection using validated synthetic data: a...

Area of Science:

• Computational biology
• Infectious disease modeling
• Machine learning applications

Background:

• Malaria is a significant global health issue necessitating advanced diagnostic tools.
• Machine learning (ML) presents opportunities for automated malaria detection.
• Limited systematic comparisons of ML algorithms using validated clinical data exist.

Purpose of the Study:

• To systematically compare five ML models for malaria detection.
• To evaluate performance using a validated synthetic dataset reflecting Sub-Saharan African conditions.
• To prioritize clinical sensitivity and cost-effectiveness in model selection.

Main Methods:

• Comparison of Naive Bayes, Logistic Regression, Random Forest, XGBoost, and Enhanced Bayesian Logistic Regression.
• Utilized a synthetic dataset with 87% representativeness against clinical benchmarks.
• Employed cost-sensitive threshold optimization, bootstrap confidence intervals, statistical significance testing, and clinical cost analysis.

Main Results:

• XGBoost demonstrated optimal performance with the highest Area Under the Curve (AUC) of 0.956 and competitive clinical cost.
• Enhanced Bayesian Logistic Regression offered comparable AUC (0.954) with interpretable coefficients.
• McNemar's test showed significant differences between XGBoost and Random Forest, but Friedman test found no overall ranking differences.

Conclusions:

• XGBoost offers an optimal balance of accuracy and cost-effectiveness for malaria screening.
• The systematic validation framework and cost-sensitive optimization provide guidance for clinical implementation.
• While synthetic data aids comparison, real-world clinical validation is crucial before deployment.