Visão computacional para auxílio na avaliação das compressões torácicas durante a RCP

• Autor(a) principal: Zankoski, João Pedro Menegas
• Data de Publicação: 2025
• Tipo de documento: Trabalho de conclusão de curso
• Idioma: por
• Título da fonte: Repositório Institucional da UTFPR (da Universidade Tecnológica Federal do Paraná (RIUT))
• Texto Completo: http://repositorio.utfpr.edu.br/jspui/handle/1/37546
• topic: Ressuscitação; Aprendizado do computador; Visão por computador; Resuscitation; Machine learning; Computer vision; CNPQ::CIENCIAS EXATAS E DA TERRA::CIENCIA DA COMPUTACAO

Cardiovascular diseases remain among the leading causes of mortality in Brazil, underscoring the importance of Cardiopulmonary Resuscitation (CPR) as a critical emergency intervention. In this context, the proper training of both healthcare professionals and laypersons is essential for increasing survival rates. This work proposes the development of a computational tool designed to support the teaching and practical assessment of CPR. The approach is based on the automated analysis of training videos, with the aim of identifying execution errors—particularly those related to chest compression rate. Such analysis enables the generation of objective, systematic feedback, thereby facilitating more precise pedagogical interventions by instructors. Additionally, we seek to implement a classification system capable of accurately distinguishing between correct and incorrect CPR performances. The main challenge of this study is the limited number of available videos, which we address through data augmentation techniques to enhance the proposed model’s generalization and robustness. We expect that the results will contribute to the improvement of CPR teaching–learning processes, with the potential to better prepare individuals for responding to cardiovascular emergencies.To this end, 16 training videos containing both correct and incorrect executions were captured; body keypoints were extracted with MediaPipe and YOLO, converted into time-series data, analysed for periodicity via the Short-Time Fourier Transform (STFT), and subsequently used to train a Long Short-Term Memory (LSTM) network optimised with Optuna. In the final tests, the classifier achieved its best performance with an F1-score of 0.95 using YOLO and 0.91 using MediaPipe, with mean precision and recall values exceeding 0.89.