PhD Oral Exam - Mohamed Elshafei, Computer Science
On the Impact and Detection of Biceps Muscle Fatigue in Wearable Sensors-Based Human Activity Recognition
This event is free
School of Graduate Studies
When studying for a doctoral degree (PhD), candidates submit a thesis that provides a critical review of the current state of knowledge of the thesis subject as well as the student’s own contributions to the subject. The distinguishing criterion of doctoral graduate research is a significant and original contribution to knowledge.
Once accepted, the candidate presents the thesis orally. This oral exam is open to the public.
Nowadays, modern sport and athletic training are very interested in wearable-based Human Activity Recognition (HAR) systems due to their cost-efficiency, portability, and convenience. However, this leads the developers to compete in developing the various HAR applications with little attention to HAR’s-related problems such as fatigue. In this thesis, we select the bicep curls as an example of a HAR activity to study the fatigue problem in wearable-based HAR. We approach the fatigue problem through three studies: first, we study the impact of fatigue in wearable-based HAR. Second, we detect the presence of fatigue during human activity, e.g., biceps curls exercise. Third, we improve the perfor- mance of fatigue detection models while reducing the test’s data consumption. Throughout our studies, we use our dataset, which consists of 3,750 repetitions of biceps curls from twenty-five volunteers between 20–46 years and with body mass index (BMI) between 24–46.
Our first study on the impact of fatigue in wearable-based HAR shows that fatigue often occurs in later sets of biceps curls. During fatigue, the completion time of later sets extends by up to 31%, while muscular endurance decreases by 4.1%. Also, our study shows that changes in data patterns often occur during fatigue, turning some features to be statistically insignificant. This can lead to a substantial decrease in performance in both subject-specific and cross-subject models. In addition, muscle fatigue can lead to various injuries such as muscle strain and tendons rupture, which may require up to 22 weeks of treatment. Therefore, it is essential to be aware of fatigue during human activity, which we address in our second study.
The second study proposes a wearable-based approach to detect fatigue in biceps curls. We provide a set of 16 most fatigue representative features from 33 extracted features. Then, we employ these features in five models to detect fatigue in biceps curls. Our study shows that a two-layer FNN achieves the highest accuracy of 98% and 88% for subject- specific and cross-subject models, respectively. We observe that the cross-subject models are preferable for a large crowd since these models can utilize crowd data. However, we observe that inter-subject data variability is usually high in the large crowd due to the physical differences among the individuals, resulting in different data patterns for the same activities. As a result, researchers may suggest using subject-specific models for each user in the crowd to achieve higher performance. Still, such a performance comes with a higher data cost of the user’s subject-specific model; therefore, improving fatigue detection in cross-subject models is essential, which is the goal of our third study.
In the third study, we propose a personalization approach as a solution to improve the cross-subject models’ performance by utilizing data from the crowd based on similarities between the test subject and users from the crowd. We extract 11 hand-crafted features to measure the similarities between the test subject and the individuals in the crowd. Then, we employ these similarities to prioritize and select the training data from the crowd for two cross-subject models. Our study shows that the personalization approach improves the performance of the cross-subject models in terms of precision by up to 7.25%, recall by up to 5.69%, accuracy by up to 6.67%, and F1-measure by up to 6.52%. Furthermore, adding 20% of the test subject’s data into the training dataset of the personalized cross-subject models can produce accurate results closer to the ones from subject-specific models.