Comparison of a Deep Learning and a Hybrid Model for Classification of an Unbalanced Urgent Cases Dataset for Human Faces

This study investigates the utilization of Convolutional Neural Networks (CNN) and a hybrid CNN-Support Vector Machine (SVM) model for classifying imbalanced emergency medical images, specifically focusing on human faces displaying nosebleeds, vomiting, and normal conditions. Developed and compared, the CNN model and hybrid CNN-SVM model demonstrated varying accuracies; the hybrid model excelled in binary classifications of normal versus abnormal images, while the CNN model performed better in other classifications. A novel dataset, created from scratch and augmented to address class imbalance, comprised 514 nosebleeds, 27 vomiting, and 514 normal images. Key metrics such as precision, recall, and F1-score were employed to evaluate model performance, revealing that the hybrid model showed superior performance in detecting abnormalities. The study underscores the necessity of further research to enhance model performance, especially in multitask classification, by utilizing larger datasets and advanced techniques. These findings lay a foundation for future research in the automatic classification of medical images and propose significant potential applications in emergency detection systems.