Date of Award
Master of Science (MS)
This thesis considers the problem of detecting eating episodes such as meals and snacks, by tracking wrist motion using smartwatch device. Previous work by our group has trained a wrist motion classifier using a large data set collected from 351 people to learn general eating behaviors. We call this a group model. This thesis investigates training the classifier with the same model architecture on new data collected by 8 people, and training the individualized classifier separately for each person. We call these individual models. The main goal in this work is to determine if individual models provide higher accuracy in detecting eating episodes, with fewer false positives, compared to the group model. By comparing their performance, we can also know if the improvement from individual models varies for each individual.
In data collection, two data sets were used. One is the individual data set, which was collected from 8 participants and each participant has at least 10 days of wrist motion 6-axis timeseries data. There are 115 days, 1,064.5 hours and 246 meals collected in total in this data set. The second one is the group data set, called Clemson All-day Data set (CAD), collected in previous work. This group data set collected from 351 participants contains 354 days, 1,133 meals, 250 eating hours and 4,680 hours in total. Two data sets were first processed using smoothing and normalization techniques and then cut along time by a sliding window to generate training and testing samples for training models.
In model training and evaluation, all models used the same convolution neural network architecture. Only one group model was trained on CAD group data set and this group model was used to compare with all other individual models. We used 5-fold cross validation to train and evaluate 5 individual models per individual. In model evaluation, we selected weighted accuracy (WAcc) as time metric to measure the models’ ability of classifying each window sample as eating or non-eating. We also selected true positive rate (TPR) and ratio of false positive over true positive (FP/TP) as episode metrics to measure model’s ability of detecting each meal episode. TPR measures how many true eating episode are detected correctly and FP/TP measures the ratio of wrong detection amount over true detection amount. Hence when TPR is larger and FP/TP is smaller, model performs better. WAcc, TPR and FP/TP were measured by cross validation.
When measuring the time metric, we found that over 8 participants, the average WAcc on all individual models is 0.819 and the average WAcc on the group model is 0.780. On average, the individual models outperform the group model. Moreover, the improvement of individual models over the group model can vary per individual. For example, in one individual data set, individual models with WAcc of 0.897 have obvious improvement compared to the group model with WAcc of 0.774. In another individual data set, WAcc of 0.958 from individual models is very close to WAcc of 0.956 from the group model.
In the measurement of episode metrics, we found that before tuning hyper-parameters Ts and Te, compared to the group model, individual models have the average improvement of 8.6% on TPR, but -14.4% on FP/TP. After tuning Ts and Te, individual models have the average improvement of 10.1% on TPR and 33.2 % on FP/TP. Tuning Ts and Te can improve the individual models’ episode metrics.
Wei, Wenkang, "Individualized Wrist Motion Models for Detecting Eating Episodes Using Deep Learning" (2021). All Theses. 3515.