Integrating In-Ear Wearable Sensors with Deep Learning for Head and Facial Movement Analysis

Accurate and inconspicuous identification of eating behavior has substantial implications for wellness surveillance, nutritional tracking, and technological assistance. Conventional approaches for detecting eating habits frequently rely on invasive or unfeasible techniques, which restrict their practicality in real-life situations. This research introduces an innovative eating recognition system that utilizes head and face motions recorded by in-ear wearing sensors and advanced deep learning algorithms. Our approach is a non-invasive and simple-to-use method that utilizes the unique patterns of head and face motions related to eating actions. We use a specially developed in-ear wearable sensor from the EarSet dataset to gather movement information while individuals are eating. The dataset comprises a range of cranial and facial motions recorded from 30 individuals. The unprocessed sensor data is subjected to pre-processing to eliminate interference and extract pertinent characteristics. We use advanced deep learning models, such as CNN, LSTM, BiL-STM, GRU, BiGRU, and our novel hybrid model CNN-LSTM, to acquire knowledge of distinctive patterns and identify eating occurrences. Our eating recognition system has been extensively tested and proven successful. The CNN-LSTM model achieved an accuracy of 95.12% and an F1-score of 95.09% during the 5-fold cross-validation studies. Our technique demonstrates superior efficiency and applicability compared to standard deep learning algorithms, as a comparative study shows. The suggested system can combine in-ear wearable sensors with deep learning to provide inconspicuous and accurate identification of eating activities. This study enhances the development of eating detection technology and introduces new opportunities for individualized wellness tracking and helpful applications. Subsequent research will prioritize the enlargement of the dataset, integration of supplementary modalities, enhancement of deep learning models for equipment with limited resources, and assessment of the long-term practicality and user reception of in-ear wearable sensors for detecting eating activities.