An emerging number of methods are proposed for eating event detection. To assess the detection performance, evaluation metrics such as precision, recall and F1 score are commonly used. However, these popular metrics only give overview of event retrieval performance and do not tell people the absolute timing error of the eating detection algorithm. Meanwhile, the eating timing error is an important marker for applications. For example, diabetic patients can manage the insulin injection time better with a preciser detected eating timing. We propose using eating timing errors, i.e. the absolute time differences at the eating start/end between an detected event and ground truth, as evaluation metrics for eating detection.
Eating spotting methodology
As an example, we estimated the eating timing errors of our eating spotting algorithm based on a one-class Support Vector Machine (ocSVM). We collected bilateral Temporalis electromyography (EMG) data from 10 participants in free-living using smart eyeglasses integrated with EMG sensors. A feature vector of only 6-dimensions were extracted and fed to the ocSVM spotter. We varied the spotting window size w with/without a fixed overlapping section length of 1 second. For the evaluation, leave-one-participant-out (LOPO) cross-validation was used.
Eating spotting results
The result showed that the highest F1 score is not always related to the lowest timing errors. The lowest timing errors (mean ± std.) were achieved at different window sizes: 21.8 ± 29.9 seconds for the eating start, and 14.7 ± 7.1 seconds for the eating end.
- Chewing monitoring
- Automatic dietary monitoring
"Free-living eating event spotting using EMG-monitoring eyeglasses", Proceedings of the 2018 IEEE EMBS International Conference on Biomedical Health Informatics (BHI '18), IEEE, March 2018.,
"Retrieval and Timing Performance of Chewing-Based Eating Event Detection in Wearable Sensors", Sensors, www.mdpi.com, 2020.,