اميرمحمد داودابادي فراهاني

Cardiovascular diseases are among the leading causes of death worldwide, an‎d early diagnosis plays a crucial role in preventing heart-related conditions. With the growing use of wearable an‎d portable ECG recording devices, automatic analysis of these signals using deep learning has become a highly active research area. However, due to limited personal data an‎d the need to protect user privacy, as well as the restricted computational power, environmental noise, an‎d small amount of personalized data on wearable devices, these models usually perform worse than those trained on large clinical datasets. To address this issue, using a central server to share knowledge among local models can help improve overall system performance. In this study, we propose a federated learning framework based on knowledge distillation for heart disease detection. The proposed base model is trained on the PTB-XL dataset to classify five different classes. Then, the outputs of five base models are combined on the server side using an XGBoost-based meta-learner to optimally integrate the knowledge from all models. During the federated learning process, each client is trained using its local data along with the output of the global model, while the server is updat‎ed through a two-level distillation process between the teacher model an‎d the clients. Experimental results show that the base model achieved an F1 score of 77.5% an‎d an AUC of 93.5%, while the proposed federated learning framework reached AUC scores of 87.35% for Client 1 an‎d 86.12% for Client 2. These findings demonstrate the effectiveness an‎d robustness of the proposed method.

بيماري‌هاي قلبي عروقي , تشخيص خودكار ECG , يادگيري عميق

Cardiovascular diseases , Automatic ECG diagnosis , Deep learning

AmirMohammad Farahani

Ali Sadr