Enhancing Chewing Detection with a CNN-BiLSTM Network Utilizing IMU and Photoplethysmography Data from an Earable Device

Detecting chewing is vital to tracking eating habits, which is necessary to preserve good well-being and health. Advancements in wearable sensors have recently enabled cap-turing a wide range of sensor data, including signals from inertial measurement units (IMUs) and photoplethysmography (PPG). This research presents a new deep-learning design that efficiently detects chewing occurrences employing IMU and PPG sensor data acquired from an early instrument. The design includes convolutional neural networks (CNN) and bidirectional long short-term memory (LSTM). The CNN-BiLSTM network presented in this study combines the advantages of both CNN and BiLSTM to capture spatial and temporal characteristics from the sensor data effectively. The CNN layers extract localized patterns and characteristics from the IMU and PPG data, whereas the BiLSTM layers represent the extended relationships and time-based changes in chewing behavior. By integrating these two formidable designs, our network attains a high level of resilience and precision in detecting chewing. We assess the efficacy of the CNN-BiLSTM network using a standardized dataset obtained from several individuals wearing wearable devices while engaging in eating events. The dataset comprises IMU and PPG sensor data that have been annotated with associated chewing sessions. The experimental findings clearly show that the suggested network outperforms the current leading approaches, obtaining an accu-racy of 98.18 % and an F1-score of 98.29 %, the highest recorded. The CNN-BiLSTM network performs better than baseline models such as CNN, LSTM, BiLSTM, GRU, and BiGRU, emphasizing the efficacy of integrating spatial and temporal modeling for chewing identification.