Abstract/Summary

The widespread adoption of wearable devices enables continuous monitoring of physiological parameters like heart rate, offering valuable insights into health. However, consumer-grade wearable data exhibits real-world noise, variations, and discontinuities across diverse populations, posing significant challenges for anomaly detection models. This paper proposes a novel deep learning approach to address these challenges, utilizing a Contractive Autoencoder (CAE) model optimized and applied specifically to noisy temporal heart rate data from wearable devices. By incorporating a contractive regularization penalty in the loss function, the model learns more robust and stable representations of the irregular data with high accuracy. Comprehensive experiments on a real-world Fitbit dataset demonstrate the proposed CAE model accurately identifies anomalous heart rate patterns missed by traditional thresholding techniques. The research encountered key challenges in ensuring model generalizability across diverse populations with natural heart rate variations, handling missing and sparse data from unreliable real-world wearable devices, and obtaining properly labelled anomaly data for robust training. Although the current model achieved promising anomaly detection results, further extensive validation on diverse datasets is essential to fully assess its capabilities across expanded demographics and use cases. Overall, this research provides an important foundation for optimizing deep learning approaches on noisy real-world wearable data through rigorous evaluation.