Intelligent Frozen Shoulder Self-Home Rehabilitation Monitoring System Jiann-I Pan* 1, Hui-Wen Chung 1, and Jen-Ju Huang 2 1 Department of Medical Informatics, Tzu-Chi University, Hua-Lien, Taiwan 2 Rehabilitation Department, Buddhist Tzu-Chi General Hospital, Hua-Lien, Taiwan * jipan@mail.tcu.edu.tw Abstract. Upper limb function rehabilitation exercises can assist in improving shoulder pain, activity, and maintain muscle strength. In addition to the regular supervision of professional rehabilitation staff, comply with the self-home rehabilitation can highlight the effectiveness. In this paper, smart phone is served as the platform which integrates an accelerometer-based senor network to facilitate the monitoring of self-home exercise rehabilitation for frozen shoulder patients. The acceleration-based sensor network consists of two developed accelerometer sensors and the smart phone build-in accelerometer, which communicated by Bluetooth protocol. The activities of upper limb exercise is recognized by the Support Vector Machine algorithms, and recorded in smart phone. The records can be used to remind the patients what he/she has done. On the other hands, the records can be uploaded to the server in hospital for helping physicians to monitoring the exercise effectiveness. The proposed approach is low cost and easy to upgrade for further different monitoring target by installing a new Android APP. Keywords: Frozen shoulder, Self-home rehabilitation, Sensor network, Smartphone, SVM. 1. Introduction Frozen shoulder, also referred as adhesive capsulitis, usually occurred after the 40- year-old and more women than men. Frozen shoulder will cause pain and restrict the range of motion (ROM) of the shoulder, and further to affect their daily lives. The early active rehabilitation treatment will help to reduce the recovery time of joint limited, and improve the daily life as eating, dressing, toilet function, and so on. Upper limb rehabilitation exercise therapy can reduce spasms and reduce pain, really effective to improve the activity of the shoulder joint, and further avoid lymphedema occurred in real life. However, the effectiveness of the rehabilitation is often unable to reveal [1]. The main reason for the low effectiveness of rehabilitation is the patients cannot really adherence the exercise prescription at home for rehabilitation. The most IST 2013, ASTL Vol. 23, pp. 265-270, 2013 SERSC 2013 265
Proceedings, The 2nd International Conference on Information Science and Technology important part of the rehabilitation is to maintain daily fixed shoulder mobility. The patients will be easy to give up according to they doubt the effectiveness of the rehabilitation. According to the MEMS technology growing quickly, the sensors that based on accelerometers and/or gyroscope are wildly used for activity assessment and recognition [1][3][4]. With the popularity of the smart mobile devices, their related applications have vigorous developed. The establishment of the intelligent mobile phone as the calculation core, combined with the accelerometer-based motion detectors, will facilitate the effectiveness of many home care activities [1][2]. In this paper, we present a wearable sensor network that integrated sensors of tri-axial accelerometer and smartphone for upper limb rehabilitation exercise monitoring system. The rehabilitation activities data are collected by accelerometers from wearable sensors and smart phone built-in, and recognized by support vector machine algorithm. In this study, there are six exercise types be monitored, i.e. (1) touching ear (external rotation), (2) fingers climbing wall- facing the wall and side to the wall, (3) pendulum clockwise and pendulum-counter clockwise, (4) front active (flexion), (5) side active (abduction), and (6) back hand raise (internal rotation). This paper is organized as follows. Section 2 introduces the proposed self-home rehabilitation exercise monitoring system architecture. Section 3 shows the experiment results. Finally, a brief conclusion is given in Section 4. 2. Self-Home Rehabilitation Exercise Monitoring System Architecture The proposed self-home rehabilitation exercise monitoring system architecture is shown in Figure 1. Fig. 1. The proposed system architecture. 266
Intelligent Frozen Shoulder Self-Home Rehabilitation Monitoring System The hardware of the proposed system includes two accelerometry-based sensors and an Android-based smartphone. The sensors responsible to collect body movement data, and the smartphone served as the main calculation unit. The main components of this system are introduced as follows. (A) Accelerometry-based sensor. The sensor is composed by the following components (1) tri-axial acceleration sensing unit LIS3LV02DQ, (2) the processing unit is MSP430F169 microcontroller, (3) a wireless transmission unit is BTM-112 Bluetooth module, (4) power supply unit use a lightweight rechargeable lithium battery. The volume of the sensor is 40mm*28mm*18mm (see Figure 2). The sampling frequency is set as 32Hz. Fig. 2. The accelerometry-based sensor. (B) Signal Filter. In order to eliminate the noise from hardware circuit, we adapted a nonlinear signal filter, i.e. median filter, to pre-process the acceleration signal. The window size is set as 7. After median filter, used low-pass (moving average) filter to smooth the acceleration signal. The window size is set as 30. (C) Segmentation. The accelerometer data were digitally filtered with a median filter to remove highfrequency noise and then segmented to isolate movement actions. The patient is asked to press the Start button when he or she beginning the periodically exercises, and to press end button when he finished for each single exercise. Each single exercise may be includes one or more actions. There is an interval of 3 seconds of time to rest between actions and action. The segmentation was performed to distinguish each action for counting how many times patient did. In order to improve the effect of rehabilitation, the patient will be required to hold stationary for more than 5 seconds when his action reached its highest ROM. The collected raw data, which includes the state of stationary (static) and the state of movement (dynamic) (see the Figure 4), are through a segmentation algorithm. There are four stages in segmentation algorithm: (1) Calculate the Energy distinguish between stationary and movement; (2) for the stationary state and the initial action, the Dynamic Time Warping (DTW) [4] is performed to distinguish different actions between segments; (3) Use Haar wavelet transform function [5] to cut out the start and end of the action (4) segment each detailed actions from the continuity of rehabilitation exercise. 267
Proceedings, The 2nd International Conference on Information Science and Technology Fig. 3. Segmenting the static and dynamic states. (D) Feature extraction. Features were derived from the accelerometer data to capture aspects of activity such as speed, smoothness, and coordination. Specifically, we estimated the following five features: (1) mean value of the accelerometer time series; (2) root mean square value of the accelerometer time series; (3) maximum value of the velocity time series; (4) minimum value of the velocity time series; and (5) entropy of the accelerometer time series (E) Classifier. In this study, the Support vector machine [6] is adapted as the core classification algorithm. The SVM classification is wildly used to mainly deal with binary data classification and regression. As there are various rehabilitation activities in the system, the SVM processing is one-versus-all, i.e. the problem is divided into N categories classification (in particular, there are six exercise types in this study). During the training phase, the collected training data is used to construct the support hyperplane. According to the identified features, the data point tagged with 1 when it approaches to one category, otherwise other categories using -1. As such, one input testing data will be tested by the N support hyperplane and then be classified into the most correctness categories. 3. Results The sensors and smartphone are wearing as Figure 4. The smartphone is placed on the wrist for easy to operate the monitoring system. Fig. 4. The sensors and smartphone are placed on the affected shoulder and the chest. Table 1 shows the number of exercises in this experiment. 268
Intelligent Frozen Shoulder Self-Home Rehabilitation Monitoring System Table 1. Exercise types and action numbers of training, testing, and evaluation. Exercise types Total actions Training Testing Validation Touching ear (see Fig.5(a)) 100 60 20 20 Fingers climbing wall (see Fig.5(b)) 200 120 40 40 Pendulum clockwise and counter clockwise 400 240 80 80 (see Fig.5(c)) front active (see Fig.5(d)) 200 120 40 40 side active (see Fig.5(e)) 200 120 40 40 Back hand raise (see Fig.5(f)) 100 60 20 20 (a)touching ear (b) fingers climbing wall (c) pendulum counter clockwise (d) front active Fig. 5. The monitored target activities of rehabilitation exercise. (e) side active (f) Back hand raise A prototype of the intelligent rehabilitation monitoring system has been developed. The experimentation is took place in the laboratory and involved three men and seven women subjects whose aged between 21 and 23. In order to simulate variety possible angles of patients happened, each subject has made the six rehabilitation exercise several times (see Table 1). The captured data of subjects S1 to S8 are mixed and divided into training group (60%) and testing group (20%). The data of subjects S9 and S10 are separated from the others which served as the validation data. The testing results and validation results are shown in Table 2. The accuracy is defined as 1 (errors / total)% where the errors include all unexpected data. Table 2. The testing result s. Exercise Type touching ear fingers climbing wall pendulum clockwise front active side active Back hand raise Accuracy testing 85 % 97.5 % 95 % 82.5 % 70 % 100 % validation 100 % 97.5 % 88.75 % 87.5 % 85 % 100 % As shows in our experiment, there are two main categories of errors that infected the final monitoring correctness. First is the segmentation error. It is easy to make a 269
Proceedings, The 2nd International Conference on Information Science and Technology counting error by jitters during maximum ROM holding states. Second is the recognizing error. It is usually caused by the similar movement. For example, the ill arm has similar movement in the front active and side active. 4. Conclusion In this paper, we have presented a low cost and expandable approach to monitoring frozen shoulder rehabilitation for patient at home. Two accelerometer-based sensors and an accelerometer build-in smartphone captured the rehabilitation exercises, and recognized by SVM classification algorithm which implemented in the smartphone. The proposed self-home rehabilitation exercise monitoring system provides three main benefits. (1) Visibility: physiatrists can follow up the exercise prescription, i.e. actually daily times of rehabilitation, by exercise records on smartphone; (2) Portability: the system is not limited to a specific location, and can carry on rehabilitation exercises anytime and anywhere; and (3) Extendibility: the main software installed in the smart phone can extend or update its functionalities through the app update procedure without modifying the hardware. Acknowledgment This project was supported by the National Science Council of Taiwan (Grant No: NSC 101-2218-E-320-003). References 1. Patel, S., Hughes, R., Hester, T., Stein, J., Akay, M., Dy, J.G., and Bonato, P., A Novel Approach to Monitor Rehabilitation Outcomes in Stroke Survivors using Wearable Technology. IEEE, Vol. 98, Issue: 3, pp.450 461, March (2010). 2. Bourke, A.K., O Brien, J.V., and Lyons, G.M., Evaluation of a threshold-based tri-axial accelerometer fall detection algorithm. Journal of Gait and Posture, Vol. 26, Issue: 2, pp.194 199, (2007). 3. Giuffrida, J.P., Lerner, A., Steiner, R., and Daly, J., Upper-extremity stroke therapy task discrimination using motion sensors and electromyography. IEEE Transactions Neural Systems and Rehabilitation Engineering, Vol. 16, Issue: 1, pp. 82 90, Feb. (2008). 4. Muscillo, R., Schmid, M., Conforto, S., and D Alessio, T., Early recognition of upper limb motor tasks through accelerometers : real-time implementation of a DTW-based algorithm. Computers in Biology and Medicine,Vol. 41, Issue: 3, pp. 164 172, (2011) 5. Haar, A., Zur theorie der orthogonalen funktionen systeme, Mathematische Annalen. Vol. 69, pp. 331 371, (1910). 6. Chang, C.C., and Lin, C.J., LIBSVM: A Library for Support Vector Machines. Journal of ACM Transactions on Intelligent Systems and Technology,Vol. 2, Issue: 3, Article 27, 27 pages,(2011). 270