Improving Classification of Sit, Stand, and Lie in a Smartphone Human Activity Recognition System Nicole A. Capela Mechanical Engineering, University of Ottawa, Ottawa, Canada Ottawa Hospital Research Institute Ottawa, Canada Edward D. Lemaire Ottawa Hospital Research Institute Ottawa, Canada Faculty of Medicine, University of Ottawa, Ottawa, Canada Natalie Baddour Dept. of Mechanical Engineering University of Ottawa Ottawa, Canada Abstract— Human Activity Recognition (HAR) allows healthcare specialists to obtain clinically useful information about a person’s mobility. When characterizing immobile states with a smartphone, HAR typically relies on phone orientation to differentiate between sit, stand, and lie. While phone orientation is effective for identifying when a person is lying down, sitting and standing can be misclassified since pelvis orientation can be similar. Therefore, training a classifier from this data is difficult. In this paper, a hierarchical classifier that includes the transition phases into and out of a sitting state is proposed to improve sit- stand classification. For evaluation, young (age 26 ± 8.9 yrs) and senior (age 73 ± 5.9yrs) participants wore a Blackberry Z10 smartphone on their right front waist and performed a continuous series of 16 activities of daily living. Z10 accelerometer and gyroscope data were processed with a custom HAR classifier that used previous state awareness and transition identification to classify immobile states. Immobile state classification results were compared with (WT) and without (WOT) transition identification and previous state awareness. The WT classifier had significantly greater sit sensitivity and F- score (p<0.05) than WOT. Stand specificity and F-score for WT were significantly greater than WOT for seniors. WT sit sensitivity was greater than WOT for the young population, though not significantly. All outcomes improved for the young population. These results indicated that examining the transition period before an immobile state can improve immobile state recognition. Sit-stand classification on a continuous daily activity data set was comparable to the current literature and was achieved without the use of computationally intensive feature spaces or classifiers. Keywords—Human activity recognition; HAR; activities of daily living; ADL; smartphone; movement; mobility; accelerometer I. INTRODUCTION Human activity recognition (HAR) using wearable sensors can offer valuable information to healthcare specialists about a person’s daily activities, thus providing insight into their mobility status and the frequency and duration of activities of daily living (ADL). High mobility classification accuracy has been achieved with systems that have multiple sensor locations; however, specialized wearable sensors are inconvenient for long term use outside a hospital setting. Single sensor systems result in higher user compliance and lower device costs, making it more practical for HAR purposes [1], [2] Smartphones are ubiquitous and contain useful sensors, and many people carry them in their daily lives, making them an ideal HAR platform. While most HAR devices can accurately distinguish when a person is active or inactive, knowing when a person is standing, sitting, or lying down can give a more representative portrait of a person’s health state. Del Rosario et al. [1] proposed that the cost of misclassifying a sitting state as standing is higher than the cost of misclassifying a sitting state as lying down, since sitting and lying are both sedentary states with lower energy expenditure. When classifying immobile states, HAR devices typically rely on phone orientation to differentiate between sit, stand, and lie [3] This is generally effective for identifying when a person is lying down, but often provides poor results for differentiating between sitting and standing, since pelvis and trunk orientation can be similar when the person is upright. For this reason, sitting and standing states are often mutually misclassified when using a single sensor location [4]. Gjoreski et al. noted that the sit-stand classification issue is largely sidestepped in the literature [5], with some researchers opting not to include both classes [6] and some studies merging the two classes together [7]. Gjoreski et al. used a context- based ensemble of classifiers and additional features designed specifically to distinguish standing and sitting in order to improve sit-stand recognition accuracy from 62% to 86%. Newly added features should be mutually uncorrelated to existing features in order to provide beneficial gains in classifier accuracy [8], which is difficult when features are derived from a single sensor or sensor location. High dimensionality feature vectors, multiple contexts, and complex classifiers can quickly become computationally intensive. Furthermore, algorithms intended for long term activity recognition must consider the trade-off between load requirements and battery life [9]. Complex models may offer only marginal improvements over simple classifiers despite significant increase in computational cost, and these improvements may disappear when the classifier is used in a real-world environment [8]. The improvements found in sophisticated models over simpler models can degrade on future samples [8], since future distributions are rarely identical to the training data distribution. A simple classifier includes the data’s most important aspects, capturing the underlying phenomena without over-fitting to the training data, and is scalable and computationally efficient [8], [10]. This project was funded by the Natural Sciences and Engineering Research Council of Canada (NSERC). Smartphones were provided by BlackBerry Inc. This full text paper was peer-reviewed at the direction of IEEE Instrumentation and Measurement Society prior to the acceptance and publication. 978-1-4799-6477-2/15/$31.00 ©2015 IEEE