Convolutional Autoencoder-Based Anomaly Detection in PPG Signals Using a Necklace Wearable Device

Detecting anomalies in PPG signals is crucial for the early identification of cardiovascular conditions, such as arrhythmias, poor perfusion, or stress induced by daily activities, thereby enabling timely interventions. This approach supports continuous, non-invasive monitoring and promotes advancements in data-driven healthcare. In this paper, we summarize the design and implementation of a wearable necklace sensor developed for monitoring the well-being of individuals during daily activities, such as driving, by acquiring PPG signals. Furthermore, we introduce an anomaly detection method for PPG signals based on a Convolutional Autoencoder (CAE). CAE architectures are particularly well-suited for tasks involving data compression and reconstruction, as they effectively capture local data relationships and preserve spatial structures. They are especially advantageous for processing PPG signals. Moreover, their reduced number of parameters compared to traditional Autoencoders (AE) makes them computationally more efficient compared to dense AE. The proposed approach is validated using a dataset of normal PPG signals, acquired through our innovative necklace sensor, for training, along with various anomalous PPG datasets sourced from public databases for testing. Experimental results demonstrate that the CAE successfully generalizes well from the training data and achieves highly effective discrimination between normal and anomalous signals, with AUC scores approaching one across all analyzed cases.