An interpretable dual-level lightweight framework for malaria diagnosis for enhanced remote diagnosis in resource-constrained settings

Certain mosquito species can transmit the deadly disease malaria to humans. During an outbreak, massive data storage and remote diagnosis are necessary. A possible solution is cloud computing, however even while it offers the possibility of remote monitoring and storage, it has a high latency. Whereas fog computing reduces latency to a greater extend and makes remote diagnosis easier. Our proposed framework with two lightweight models for diagnosis of malaria has addressed these issues. The level-1 diagnosis make use of Optimized Lightweight MLP. On the basis of the symptoms, the most likely cases are identified by the model and the result is forwarded to the healthcare provider who will serve as the intermediary. The proposed model in phase 1 has been combined with Anchor, an explainable AI (XAI), to aid in the diagnosis. The suggested model outperformed current models in terms of efficiency, achieving an impressive 96% accuracy rate. Strong model performance is indicated by the training and test curve’s good AUC (Area Under the Curve) of 0.99. A lightweight customized CNN model with GradCAM assists in finding confirmed cases in the second level. Comparatively, our proposed lite-weight CNN model showed greater efficiency with fewer parameters, a smaller model size and equivalent or more accuracy of 95% achieving a good balance between computational efficiency and performance. Data augmentation and Image processing techniques improved the efficiency of the model. Finally, a simulation that mimics the operations of the proposed framework has been carried out using iFogSim.