Aplicação móvel para detecção de assimetria facial em pacientes com AVC utilizando redes neurais convolucionais

Abstract

This monograph presents the development of a mobile application based on deep learning to support the early detection of stroke through facial asymmetry analysis. The study aims to investigate the potential of Convolutional Neural Networks to identify facial patterns associated with motor dysfunction, contributing to the initial recognition of stroke symptoms in out-of-hospital settings. The adopted methodology involves the use of the Toronto NeuroFace dataset, which was subjected to preprocessing and data augmentation techniques in order to normalize the images and increase training variability. A customized CNN architecture was designed and evaluated using statistical metrics such as accuracy, sensitivity, specificity, F1-score, and confusion matrix. In parallel, a mobile application was developed using the Flet framework, enabling facial image upload and visualization of inference results. The obtained results demonstrate satisfactory performance of the proposed model, with emphasis on high recall values, indicating a strong ability to identify positive cases. The validation of the functional prototype confirms the technical feasibility of integrating the trained model into a mobile application, although limitations related to internet dependency, the use of static images, and dataset generalization were identified. It is concluded that the proposed solution shows potential as a support tool for early stroke screening, with future improvements suggested through on-device processing, real-time video analysis, and integration with telehealth systems.