Retrieval and Timing Performance of Chewing-Based Eating Event Detection in Wearable Sensors

Abstract

We present an eating detection algorithm for wearable sensors based on first detecting chewing cycles and subsequently estimating eating phases. We term the corresponding algorithm class as a bottom-up approach. We evaluated the algorithm using electromyographic (EMG) recordings from diet-monitoring eyeglasses in free-living and compared the bottom-up approach against two top-down algorithms. We show that the F1 score was no longer the primary relevant evaluation metric when retrieval rates exceeded approx. 90%. Instead, detection timing errors provided more important insight into detection performance. In 122 hours of free-living EMG data from 10 participants, a total of 44 eating occasions were detected, with a maximum F1 score of 99.2%. Average detection timing errors of the bottom-up algorithm were 2.4 ± 0.4 s and 4.3 ± 0.4 s for the start and end of eating occasions, respectively. Our bottom-up algorithm has the potential to work with different wearable sensors that provide chewing cycle data. We suggest that the research community report timing errors (e.g., using the metrics described in this work).

EatingTrak

“…Using glasses equipped with multiple sensors is another choice. Glasses equipped with EMG [58][59][60][61], load cell [14], piezoelectric sensor [18]. Multi-modal sensors on glasses [5,38] and necklaces [64] are also used.…”

“…EatingTrak was developed based on the observation that one key challenge of identifying eating moments from miscellaneous body movements in the wild using a wrist-mounted IMU (acceleration and angular velocity), compared to other systems using customized hardware [5,9,30,60,64], is the lack of enough contextual information of body posture. As a result, the eating gesture (raising the wrist towards the mouth) can be highly similar to other wrist movements in daily activities, if only examining the movement on the wrist [17,46,48].…”

“…To free users from tedious self-reporting tasks, a variety of wearable technologies were invented to automatically identify eating moments in the wild, which is the first step towards a fully automated eating activity journaling system. However, the existing solutions have two limitations: 1) They either require using a new hardware, which may not immediately applicable [5,9,30,60,64]; 2) The ones using commodity wearables (e.g., smartwatch) can not recognize eating moments in high time resolution. For instance, Thomaz et al [48] used a wrist-mounted IMU to recognize eating moments in the wild in time resolutions of 5 minutes, 30 minutes, 60 minutes with F1 scores of around 30%, 40%, and 76.1% respectively.…”

Audio-Based Onset Detection applied to Chewing Cycle Segmentation

“…automated dietary monitoring (ADM) [3]. For example, recognised chewing cycles can be used to estimate meal and snack times [14]. While several sensor technologies have been proposed to detect chewing cycles, e.g.…”

Special Issue “Advanced Signal Processing in Wearable Sensors for Health Monitoring”

“…Physical health monitoring using wearable sensors is presented in two papers related to physical movements, such as the prediction of joint momentum for the purpose of predicting the force generated by the muscle using an ANN for the purpose of skeleton control [ 11 ]. Another related work presented in [ 12 ] is based on the detection of chewing event using EMG signals.…”