Privacy-preserving ECG data collection for arrhythmia classification

초록

ECG is important data used for diagnosing arrhythmias, which are dangerous heart conditions. Recently, deep learning techniques have made it possible to classify arrhythmias from ECG data with high accuracy. Thus, large amounts of ECG data must be collected to train deep learning models. However, ECG data can be used to identify individuals, which can lead to privacy breaches if the data collector abuses it or if data leakage occurs. Therefore, a method is needed that prevents personal identification from ECG data while still enabling arrhythmia classification. In this paper, we propose two deep learning models that utilize the attention mechanism. The first model extracts features from ECG data and determines which features are important for identification and classification, respectively. Then, the second model adds stronger noise to features important for identification and weaker noise to features important for classification. This noise disrupts personal identification without hindering arrhythmia classification. In experiments using the MIT-BIH Arrhythmia Database, we demonstrate that the proposed method maintains the arrhythmia classification accuracy of 95.06% while effectively preserving privacy by reducing the personal identification accuracy to 2.68%. Additionally, despite the addition of noise, high classification performance is maintained across various metrics. Therefore, collecting ECG data processed with the proposed method can preserve privacy while still enabling arrhythmia classification.

키워드