Regresar

Active Learning Approach for Clinical Noise Characterization in Long-Term Electrocardiogram Monitoring

Abstract:

Long-term monitoring (LTM) of electrocardiogram (ECG) can help identify intermittent arrhythmias that may not appear on shorter recordings, but it is often contaminated by noise, affecting their diagnostic utility. Recognizing clinically valid ECG parts can improve the performance of signal processing systems and reduce cardiologists’ analysis time. This study computes the distortion of the ECG based on clinical readability rather than quantitative measures. Although we have proposed various machine learning systems to characterize clinical noise, we must expand our limited data to enhance performance. The main objective is to develop an active learning (AL) methodology to increase labeled data and improve our clinical noise classification models. Our experimental database comprises 8.467 excerpts of 5 seconds, labeled by a trained expert and a cardiologist, balanced across both clean and noisy categories. We adopt an AL scheme with a 1-D Convolutional Neural Network (CNN) based on an autoencoder, which iteratively incorporates new examples into the training set based on their class probability. The F1-score test curves illustrate that the AL scheme outperforms sample selection at random. This approach can refine the models by increasing labeled data for training, thus enhancing health professionals’ confidence in medical decision support systems for clinical practice applications.