Sazonova N, Browning R, Sazonov E, Proceeding of The 2nd International Conference on Ambulatory Monitoring of Physical Activity and Movement, Glasgow, Scotland, May 2011.
Use of wearable sensors coupled with the processing and visualization power of mobile communication devices (smart phones) may be an attractive and convenient way to monitor and to provide real-time feedback about a person’s levels of physical activity and Energy Expenditure (EE) that would help in achieving and maintaining a healthy lifestyle . In this paper we propose an integration of a shoe-based wearable sensor system with a cell phone for real-time prediction and display of time spent in various postures/physical activities and the resulting EE. To deal with processing power and memory limitations of the cell phone, we introduce new low-power algorithms using multinomial logistic regression for posture and activity classification and EE prediction that allow real-time execution on a wearable platform.
The study involved seventeen subjects (weight = 70.5 ± 15.8 kg, BMI = 25.2 ± 6.5) wearing a sensor system integrated into the shoe and wirelessly connected to a smart phone. The shoe system included five insole pressure sensors and a heel-mounted three-dimensional accelerometer. Subjects performed fifteen different everyday activities inside of a room calorimeter (average visit duration 3.8±0.36hours). Several predictor metrics (mean, standard deviation and entropy) were derived from each sensor signal. A multinomial logistic regression model for classification of every 2-sec interval into one of four posture/activity groups (“Sit”, ”Stand”, ”Walk”, ”Cycle”) and an EE prediction model branched by activity were developed and validated using leave-one-out approach.
The results suggest that the proposed wearable sensor system and algorithms may be capable of accurate posture and activity classification (84±8% average per-minute accuracy in classification of 4 major postures/activities) and EE predictions (total EE error of 4.55%, RMSE of 0.54 METs). Use of logistic regression in place of our previously developed support vector machine prediction algorithms  reduced the running time by a factor of 190 (from from 655ms to 3.5ms) and reduced the memory needed for model storage by a factor of 104. Implementing 2-second EE predictions (compared to commonly used 1-minute intervals ) further reduced memory requirements for sensor signal storage by a factor of 30 while only slightly increasing the total error (from 4.55% to 6.18%) with other performance characteristics practically unchanged.
DISCUSSION AND CONCLUSION
The proposed methodologies for automatic posture and activity classification and EE prediction provide accurate, fast and computationally feasible approach that is directly applicable for the integration into a wearable cell phone-based system for physical activity monitoring and real-time biofeedback.
 E.S. Sazonov, G. Fulk, J. Hill, Y. Schutz, and R. Browning, “Monitoring of Posture Allocations and Activities by a Shoe-Based Wearable Sensor,” Biomedical Engineering, IEEE Transactions on, vol. 58, 2011, pp. 983-990.
 N. Sazonova, R. Browning, and E. Sazonov, “Accurate Prediction of Energy Expenditure Using a Shoe-Based Activity Monitor,” Medicine & Science in Sports & Exercise, vol. 43, Jul. 2011, pp. 1312-21.