A Disentangled VAE-BiLSTM Model for Heart Rate Anomaly Detection.

Area of Science:

• Artificial Intelligence
• Cardiovascular Health
• Wearable Technology

Background:

• Cardiovascular diseases (CVDs) are a leading global cause of death, with risk factors including lifestyle choices and conditions like diabetes.
• Wearable devices collect physiological data, such as heart rate, offering potential for CVD monitoring and prevention.
• Current wearable data often lacks immediate user comprehension of health risks, necessitating advanced analytics.

Purpose of the Study:

• To develop an unsupervised anomaly detection model for heart rate data from wearable devices.
• To enhance the interpretation of wearable-collected physiological data for early CVD risk identification.
• To provide users with more accessible and actionable health insights from their wearable devices.

Main Methods:

• Utilized a disentangled variational autoencoder (β-VAE) with a bidirectional long short-term memory (BiLSTM) network.
• Applied the model to unsupervised anomaly detection in heart rate data collected during sleep from eight participants.
• Tested the model using mean heart rate data sampled at 30-second and 1-minute intervals.

Main Results:

• The proposed β-VAE with BiLSTM backend demonstrated superior performance in detecting heart rate anomalies compared to other algorithms.
• The model's effectiveness was consistent across different participants and data sampling intervals.
• Outperformed established anomaly detection methods in most tested scenarios.

Conclusions:

• The developed AI model effectively detects anomalies in sleep heart rate data from wearables.
• Integration of such anomaly detection algorithms can significantly improve the utility of wearable devices for health monitoring.
• Wearable technology, enhanced with AI, can offer users clearer insights into potential cardiovascular risks.