Anomaly Detection In Medical Data Streams With Sustainable Unsupervised Learnin

The research on anomaly detection based on medical data streams aims to detect abnormal signals in the medical data flow.By using intelligent algorithms for anomaly detection in medical data streams,the efficiency of clinical diagnosis can be effectively improved.This paper focuses on two problems:(1)in the identity recognition problem,the electrocardiogram(ECG)data of other users are considered abnormal relative to a specific user,and(2)in the arrhythmia detection problem,ECG data containing arrhythmias are considered abnormal compared to normal ECG data.Currently,mainstream methods for addressing these two problems tend to train a comprehensive classification model for all users,without considering privacy risks,class imbalance issues,personalized models,and real-time anomaly detection scenarios.To address these problems,this paper conducts the following research:For the identity recognition problem based on ECG data streams: This paper proposes an identity recognition system called PerAE based on an autoencoder.PerAE trains an autoencoder called Attention-MemAE for each user for identity recognition.Attention-MemAE has high privacy and does not suffer from class imbalance issues in the training data.Attention-MemAE includes attention mechanism and memory module,which can effectively extract personalized features.This paper also introduces a dynamic updating mechanism,allowing Attention-MemAE to be deployed in real-time operational scenarios.For the arrhythmia detection problem based on ECG data streams: This paper proposes an arrhythmia detection system called PerAD based on an autoencoder.PerAD trains a lightweight autoencoder called ShaAE for each user.ShaAE has fewer parameters and faster inference speed,while avoiding high privacy risks and class imbalance issues.Similarly,this paper introduces a dynamic updating mechanism to enable real-time operation of ShaAE.The paper also utilizes a variational autoencoder for fast inference to generate simulated data for auxiliary training of ShaAE.