A Hybrid Machine Learning Model for Malaria Prediction and Classification

Abstract

Malaria remains a significant global health challenge, particularly in sub-Saharan Africa, where it accounts for 96% of malaria-related deaths worldwide. One of the critical failures in malaria prevention is the lack of efficient diagnostic tools. This study addresses this gap by developing and validating a machine learning algorithm to detect and classify malaria parasites in blood samples. A hybrid machine learning model was developed in Python using the OpenCV, Keras, and TensorFlow packages. The model used a VGG-19 architecture with transfer learning and data augmentation. For training, testing, and validation, 2,207 microscopic images of blood samples representing severe (complicated) malaria, mild (uncomplicated) malaria, and non-malarial infections were obtained from the National Institute of Health's (NIH) official database. The dataset was split into training (60%), validation (20%), and testing (20%) sets. Stain normalization, label encoding, and image preprocessing were performed to optimize model performance. The model achieved 95% accuracy during training, increasing to 96.24% during implementation and testing. Stratified five-fold cross-validation yielded a mean accuracy of 95.87% ± 0.83%, confirming robustness across different data partitions. External validation on an independent dataset achieved 96.24% accuracy with an AUC-ROC of 0.985. Comparative benchmarking demonstrated that the proposed VGG-based hybrid model outperformed alternative architectures, including ResNet50, DenseNet121, Xception, and classical machine learning approaches. The model successfully characterized various stages of the Plasmodium parasite life cycle, including trophozoites and gametocytes, with high sensitivity and specificity. The developed hybrid machine learning model offers a promising alternative to conventional microscopic diagnosis, with improved accuracy, reduced diagnostic time, and reduced reliance on highly skilled personnel. This tool has significant potential for deployment in malaria-endemic regions, particularly in resource-limited settings.