Development of an a wearable physical activity monitoring system with an adaptive pattern-recognition algorithm
Abstract
Measurement and classification of physical activity in terms of energy expenditure in free-living individuals represents a methodological challenge in epidemiologic research. Body-worn motion sensors such as accelerometers, gyroscope and earth magnetic field sensors make it possible, to collect data that contain information about the underlying physical activities. With the help of pattern recognition techniques this information can be extracted for objective physical activity monitoring. On the other hand heart rate monitors are closely linked to energy expenditure and provide good estimates at high levels of activity.
In this study the use of accelerometers, gyroscopes, earth magnetic field sensors and heart rate monitors for physical activity recognition and energy expenditure estimation are investigated. In particular we were interested in the detection of 7 activity classes, that have similar relationships between heart rate, movement pattern and energy expenditure. Those 7 classes were: resting, standing work, sitting work, walking, biking, running and non-cyclic activities typically in sports activities. We were interested in determining the type and placement of sensors that optimize the recognition of the 7 classes, the effect of low pass filtering and the length of the sliding window used in feature extraction on recognition rate and the factors that significantly contribute to the prediction of energy expenditure.
31 Subjects (15 Women, 16 Men, Age: 38 ± 13 y, height: 1.73 ± 0.07 m, weight: 70.65 ± 12.04) had to perform several activities out of the 7 activity classes mentioned above. Datasets were recorded using 3-dimensional accelerometer, gyroscopes and earth magnetic field sensors on 9 different parts of the body (arms, wrists, knee, ankles and hip), with a frequency rate of 100 Hz. Simple features such as mean, signal energy, signal variance and frequency entropy were extracted and their ability to discriminate between the 7 activity classes were quantified using the estimated error probability of are minimum error rate classifier using normalized histograms of feature values conditioned by the classes. The features were ranked and scores were assigned to the best discriminating features in order to determine which sensor type and placement optimizes the discrimination of the activity classes. Features were then extracted again using differently filtered signals and different window sizes and presented to a maximum likelihood classifier. For each activity class forward and backward variable selection for regression analysis was used to determine the significant factors in a linear model that predicts energy expenditure from the subjects profile information, heart rate and signal energy of the accelerometer data.
Of all the scores that were assigned to the best discriminating features, 60% were assigned to features that were extracted from signals recorded from the legs, 31% to features that were extracted from the arms and 7 % to features extract from the hip. 2 % of the scores were assigned to heart rate. 74% of the scores were assigned to features extracted from accelerometer signal. Much less scores were assigned to features extracted from gyroscopes, earth magnetic field sensors and heart rate (16%, 8%, 2% respectively)
For low intensity activities such as resting and sitting activities as well as for the most intensive activity class (running) recognition rate is independent of the window size that was used to extract the features. For running and resting the choice of the filter cut-off is also irrelevant. For sitting activities it is an advantage to let higher frequencies remain in the signal. Recognition rate increases with longer window sizes for standing activities, biking, and sport/games, so does letting only lower frequencies in the signal for the same activity classes.
Heart rate was a highly significant factor in the linear model for the prediction of energy expenditure in standing work, walking, biking, sport/games, running (for all: p < 0.01). Accelerometer data was only a significant factor in the model for the prediction of energy expenditure in standing work, walking and running (for all: p < 0.04).
