온디바이스 심혈관 진단을 위한 저복잡도의 재구성형 다단계 이진 분류

Abstract

This paper proposes an efficient and reconfigurable deep neural network architecture for multi-label classification of cardiovascular diseases (CVDs) based on electrocardiogram (ECG) signals for on-device appli- cation. It employs conditional and iterative multi-stage binary classifiers to replace the conventional single-stage multi-label classifier that can’t simply reconfigure structure without redesigning hardware such as GPUs or tensor processing cores if the classifier needs to support more labels. The proposed architecture preprocesses multi- labeled disease data into a binary classification-friendly format to train the binary classifier in each stage and com- putes the optimal decision thresholds at each stage using the Youden index. When classifying diseases, only one binary classifier will be iteratively reused by loading different neural network model parameters for each disease classification. In this architecture, new disease classification can be easily implemented by saving new parameters in the system memory and increasing the iteration number programmably without redesigning the hardware. Based on the evaluation with PTB-XL-ECG dataset, the proposed model demonstrated that it can achieve the better Hamming Loss (0.1365) and macro AUC (0.9255) as the ResNet50-based multi-label classifier with over 1000x fewer parameters and scalability. In addition, we confirm that it can further reduce the Hamming Loss by 0.006 compared to identical AUC multi-binary classifiers by placing normal signal classification at the first stage and skipping dis- ease classifications stage