Detecção de inversão de eletrodos no ECG de 12 derivações usando Deep Learning: uma comparação de desempenho com arquiteturas CNN e RNN

Abstract

The telecardiology service of the Telehealth Network of Minas Gerais (TNMG) begins with the acquisition of a 12-lead electrocardiogram (ECG) in a remote unit, which is then sent to Belo Horizonte, where a specialist physician performs the interpretation. TNMG issues more than 4,000 ECG reports per day; however, one of the main challenges is electrode misplacement, which compromises the quality of the exam and may lead to diagnostic errors. Although there are algorithms in the literature capable of detecting this type of error, many do not account for multiple forms of electrode reversal, do not employ modern deep learning techniques, or have not been validated on datasets representative of real-world clinical settings. Additionally, they often present limitations in terms of computational efficiency. Given this context, this study aims to investigate and compare the performance of different deep learning architectures in the automatic detection of the five most common electrode misplacements observed in clinical practice during ECG acquisition, using a dataset derived from a real telecardiology system, TNMG. To achieve this, four neural network models were developed: CNN, CNN+RNN, CNN+BiLSTM, and CNN+BiGRU. All models were trained using a dataset composed of 22,886 ECG exams, divided into ECGTrainingSet (70%) for training and ECGTestSet (30%) for testing. The exams were preprocessed to improve signal quality, and the electrode reversals were applied through computational manipulation. The models were evaluated based on their ability to correctly classify ECG types using the metrics Sensitivity (Se), Specificity (Sp), Positive Predictive Value (PPV), and F2-score, as well as computational metrics such as the number of parameters, memory usage, and GFLOPS. The CNN+RNN and CNN+BiGRU models demonstrated the best balance between performance and efficiency, achieving Se ≥ 94% and F2 ≥ 96% across multiple classes. The results of this study have the potential to reduce the need for repeat exams caused by acquisition errors, while also contributing to the reliability of telecardiology systems by offering an intelligent tool for the automatic detection of electrode misplacements, optimizing TNMG’s operations and increasing patient safety.