An Enhanced Autoencoder-Based Anomaly Detection Model for Time Series Data From Wearable Medical Devices

Area of Science:

• Biomedical engineering
• Data science
• Artificial intelligence

Background:

• Smart wearable devices are vital for real-time health monitoring using multi-featured time series data.
• Analyzing this data is crucial for early disease detection and timely treatment.
• Robust anomaly detection models are needed for effective feature learning and diagnosis.

Purpose of the Study:

• To propose an enhanced autoencoder-based anomaly detection model for time series data from wearable medical devices.
• To improve the accuracy and efficiency of early health diagnosis through advanced data analysis.

Main Methods:

• Utilized a convolutional neural network (CNN) for multi-feature correlation learning.
• Employed a long and short-term memory (LSTM) network to capture sequence correlations.
• Incorporated a multi-head attention mechanism to handle long sequences and residual loss to mitigate vanishing gradients.

Main Results:

• Achieved high accuracy: 95.37% on the HeartDisease dataset and 95.56% on the MIMIC dataset.
• Outperformed baseline methods by 8.6% and 12.3% on the respective datasets.
• Demonstrated efficient sequential data analysis and effective capture of long-term dependencies.

Conclusions:

• The proposed model significantly enhances the success rate of early health diagnosis.
• It offers efficient analysis of sequential data from wearable medical devices.
• The model shows strong potential for improving patient outcomes through timely interventions.