Filtering by
- Language: English
Human activity recognition is the task of identifying a person’s movement from sensors in a wearable device, such as a smartphone, smartwatch, or a medical-grade device. A great method for this task is machine learning, which is the study of algorithms that learn and improve on their own with the help of massive amounts of useful data. These classification models can accurately classify activities with the time-series data from accelerometers and gyroscopes. A significant way to improve the accuracy of these machine learning models is preprocessing the data, essentially augmenting data to make the identification of each activity, or class, easier for the model. <br/>On this topic, this paper explains the design of SigNorm, a new web application which lets users conveniently transform time-series data and view the effects of those transformations in a code-free, browser-based user interface. The second and final section explains my take on a human activity recognition problem, which involves comparing a preprocessed dataset to an un-augmented one, and comparing the differences in accuracy using a one-dimensional convolutional neural network to make classifications.
The purpose of this study is to determine the feasibility of three widely used wearable sensors in research settings for 24 h monitoring of sleep, sedentary, and active behaviors in middle-aged women.
Methods
Participants were 21 inactive, overweight (M Body Mass Index (BMI) = 29.27 ± 7.43) women, 30 to 64 years (M = 45.31 ± 9.67). Women were instructed to wear each sensor on the non-dominant hip (ActiGraph GT3X+), wrist (GENEActiv), or upper arm (BodyMedia SenseWear Mini) for 24 h/day and record daily wake and bed times for one week over the course of three consecutive weeks. Women received feedback about their daily physical activity and sleep behaviors. Feasibility (i.e., acceptability and demand) was measured using surveys, interviews, and wear time.
Results
Women felt the GENEActiv (94.7 %) and SenseWear Mini (90.0 %) were easier to wear and preferred the placement (68.4, 80 % respectively) as compared to the ActiGraph (42.9, 47.6 % respectively). Mean wear time on valid days was similar across sensors (ActiGraph: M = 918.8 ± 115.0 min; GENEActiv: M = 949.3 ± 86.6; SenseWear: M = 928.0 ± 101.8) and well above other studies using wake time only protocols. Informational feedback was the biggest motivator, while appearance, comfort, and inconvenience were the biggest barriers to wearing sensors. Wear time was valid on 93.9 % (ActiGraph), 100 % (GENEActiv), and 95.2 % (SenseWear) of eligible days. 61.9, 95.2, and 71.4 % of participants had seven valid days of data for the ActiGraph, GENEActiv, and SenseWear, respectively.
Conclusion
Twenty-four hour monitoring over seven consecutive days is a feasible approach in middle-aged women. Researchers should consider participant acceptability and demand, in addition to validity and reliability, when choosing a wearable sensor. More research is needed across populations and study designs.
Validity of the Rapid Eating Assessment for Patients for assessing dietary patterns in NCAA athletes
Athletes may be at risk for developing adverse health outcomes due to poor eating behaviors during college. Due to the complex nature of the diet, it is difficult to include or exclude individual food items and specific food groups from the diet. Eating behaviors may better characterize the complex interactions between individual food items and specific food groups. The purpose was to examine the Rapid Eating Assessment for Patients survey (REAP) as a valid tool for analyzing eating behaviors of NCAA Division-I male and female athletes using pattern identification. Also, to investigate the relationships between derived eating behavior patterns and body mass index (BMI) and waist circumference (WC) while stratifying by sex and aesthetic nature of the sport.
Methods
Two independent samples of male (n = 86; n = 139) and female (n = 64; n = 102) collegiate athletes completed the REAP in June-August 2011 (n = 150) and June-August 2012 (n = 241). Principal component analysis (PCA) determined possible factors using wave-1 athletes. Exploratory (EFA) and confirmatory factor analyses (CFA) determined factors accounting for error and confirmed model fit in wave-2 athletes. Wave-2 athletes' BMI and WC were recorded during a physical exam and sport participation determined classification in aesthetic and non-aesthetic sport. Mean differences in eating behavior pattern score were explored. Regression models examined interactions between pattern scores, participation in aesthetic or non-aesthetic sport, and BMI and waist circumference controlling for age and race.
Results
A 5-factor PCA solution accounting for 60.3% of sample variance determined fourteen questions for EFA and CFA. A confirmed solution revealed patterns of Desserts, Healthy food, Meats, High-fat food, and Dairy. Pattern score (mean ± SE) differences were found, as non-aesthetic sport males had a higher (better) Dessert score than aesthetic sport males (2.16 ± 0.07 vs. 1.93 ± 0.11). Female aesthetic athletes had a higher score compared to non-aesthetic female athletes for the Dessert (2.11 ± 0.11 vs. 1.88 ± 0.08), Meat (1.95 ± 0.10 vs. 1.72 ± 0.07), High-fat food (1.70 ± 0.08 vs. 1.46 ± 0.06), and Dairy (1.70 ± 0.11 vs. 1.43 ± 0.07) patterns.
Conclusions
REAP is a construct valid tool to assess dietary patterns in college athletes. In light of varying dietary patterns, college athletes should be evaluated for healthful and unhealthful eating behaviors.
The coffee cup system can be simplified and modeled by a cart-and-pendulum system. Bazzi et al. and Maurice et al. present two different cart-and-pendulum systems to represent the coffee cup system [1],[2]. The purpose of this project was to build upon these systems and to gain a better understanding of the coffee cup system and to determine where chaos existed within the system. The honors thesis team first worked with their senior design group to develop a mathematical model for the cart-and-pendulum system based on the Bazzi and Maurice papers [1],[2]. This system was analyzed and then built upon by the honors thesis team to build a cart-and-two-pendulum model to represent the coffee cup system more accurately.
Analysis of the single pendulum model showed that there exists a low frequency region where the pendulum and the cart remain in phase with each other and a high frequency region where the cart and pendulum have a π phase difference between them. The transition point of the low and high frequency region is determined by the resonant frequency of the pendulum. The analysis of the two-pendulum system also confirmed this result and revealed that differences in length between the pendulum cause the pendulums to transition to the high frequency regions at separate frequency. The pendulums have different resonance frequencies and transition into the high frequency region based on their own resonant frequency. This causes a range of frequencies where the pendulums are out of phase from each other. After both pendulums have transitioned, they remain in phase with each other and out of phase from the cart.
However, if the length of the pendulum is decreased too much, the system starts to exhibit chaotic behavior. The short pendulum starts to act in a chaotic manner and the phase relationship between the pendulums and the carts is no longer maintained. Since the pendulum length represents the distance between the particle of coffee and the top of the cup, this implies that coffee near the top of the cup would cause the system to act chaotically. Further analysis would be needed to determine the reason why the length affects the system in this way.
protocols, including within sleep-focused studies. This study seeks to address accuracy of
accelerometer data in detection of the beginnings and ends of sleep bouts in young adults with
polysomnography (PSG) corroboration. An existing algorithm used to differentiate valid/invalid wear
time and detect bouts of sleep has been modified with the goal of maximizing accuracy of sleep bout
detection. Methods: Three key decisions and thresholds of the algorithm have been modified with three
experimental values each being tested. The main experimental variable Sleepwindow controls the
amount of time before and after a determined bout of sleep that is searched for additional sedentary
time to incorporate and consider part of the same sleep bout. Results were compared to PSG and sleep
diary data for absolute agreement of sleep bout start time (START), end time (END) and time in bed
(TIB). Adjustments were made for outliers as well as sleep latency, snooze time, and the sum of both.
Results: Only adjustments made to a sleep window variable yielded altered results. Between a 5-, 15-,
and 30-minute window, a 15-minute window incurred the least error and most agreement to
comparisons for START, while a 5-minute window was best for END and TIB. Discussion: Contrary
to expectation, corrections for snooze, latency, and both did not substantially improve agreement to
PSG. Algorithm-derived estimates of START and END always fell after sleep diary and PSG both,
suggesting either participants’ sedentary behavior beginning and ends were at a delay from sleep and
wake times, or the algorithm estimates consistently later times than appropriate. The inclusion of a
sleep window variable yields substantial variety in results. A 15-minute window appears best at
determining START while a 5-minute window appears best for END and TIB. Further investigation on
the optimal window length per demographic and condition is required.