Evaluación del riesgo de arritmia maligna basada en mediciones de ECG móviles de derivación I utilizando aprendizaje...

This paper presents an algorithm capable of estimating malignant arrhythmia risk based on a short lead-I ECG record. We chose the mean and relative standard deviation of beat-to-beat QRST integrals as feature parameters. The algorithm was trained on a learning set consisting of three subgroups: 55 healthy subjects, 48 patients without malignant arrhythmia history, and 48 malignant arrhythmia patients. These subgroups represented the normal, moderate, and high risk, respectively. The 3 nearest neighbors (3-NN), and the 1-D and 2-D Bayesian classifiers were used as supervised machine-learning techniques. The test set contained ECG signals of 63 healthy subjects and 52 patients with confirmed malignant arrhythmia records. We obtained the best classification results with the 2-D Bayesian classifier, which produced a decision efficiency of 87.30% and 94.23% for the normal and malignant arrhythmia cases, respectively. Slightly lower results were achieved by the 3-NN method (80.95%, 94.23%) and the 1-D Bayesian classifier (77.78%, 94.23%). Considering the QTc parameter instead of the QRST integral produced a relatively low decision efficiency in the malignant arrhythmia case (84.62%). The proposed method performs best with the 2-D Bayesian method, while it is still efficient with the 3-NN classifier. According to our current knowledge, our algorithm is the first one which only requires a single-channel ECG as input and efficiently estimates malignant arrhythmia risk at the same time. As the proposed method relies only on lead-I ECG, it can be very useful in mobile ECG systems (e.g., in WIWE), making out-of-hospital risk assessment possible.