Interpretable intelligent diagnosis for Lower Respiratory Tract Infections(LRTIs)

Abstract

Developing countries including Uganda face a major public health challenge from the Lower respiratory tract Infections(LRTIs), mainly common within the vulnerable groups encircling children, elderly and immuno-compromised individuals. Pneumonia alone is said to account for ten percent of under-five deaths in Uganda annually. Tuberculosis and bronchitis, are also being common LRTIs, mainly caused by numerous pathogens like fungi, bacteria,and viruses and according to World Health Organization(WHO), over 3 million deaths occur yearly on the globe and this is due to late diagnosis being a major contributor to this high mortality rate. In the last five years Machine Learning (ML) has shown positiveresults for detection and diagnosing LRTIs including pnemonia and Tuberclosis, since ML models look at internal model parameters , use a parallel data point through interchanging some of the features for which the predicted outcome changes in a relevant wa. However, even the existing machine learning models for LRTIs detection often lack transparency in their resolution, making it difficult for patients to understand and trust their results.This study empathized interpretable deep learning model for Lower Respiratory Tract Infection (LRTIs) inorder to give patients and health experts understanding and clarity. Basically, to achieve this goal, transfer learning technique was used with different frameworks based on pretrained models VGG19, CNN, Vision Transformers, Mobilenet,Desnet and Resnet50 to detect LRTIs. Confusion matrix, Accuracy and validation were used for algorithms performance.further more, we examined the explainable AI techniques basically GRAD-CAM, Saliency, overlay, GRAD-CAM++ and Score-CAM. Therefore, we demonstrated the effectiveness of the developed interpretable ML model along with VGG19 performing excellent at an accuracy of 97.12% , validation 90.54% respectively and outperforming the other existing state of art models for LRTIs detection while promoting responsible AI practices for LRTIs diagnosis in-order for reduction in the mortality rate of o