You are on page 1of 4

A Wireless Sensor System for Motion Analysis of Parkinsons Disease Patients

Geoffrey Lo1, Ashwin Ram Suresh2, Leo Stocco1, Sergio Gonzlez-Valenzuela1, and Victor C. M. Leung1
1

Department of Electrical and Computer Engineering The University of British Columbia 2332 Main Mall, Vancouver BC, Canada V6T1Z4 {glo,leos,sergiog,vleung}@ece.ubc.ca Department of Electrical and Computer Engineering Indian Institute of Technology - Madras Chennai-600036, Tamil Nadu, India ashwinram2111@gmail.com

Abstract We describe the initial design, implementation and testing of a wearable sensor system employed for human motion analysis. Our proposed system is part of an ongoing investigation aimed at efficiently timing the self-administration of prescription drugs in Parkinsons disease patients by using wireless sensors to capture distinctive motion patterns that indicate the onset of dyskinesia lapses as the prescription drug wears off. Our prototype incorporates 3 pairs of accelerometer/gyroscope sensors, each connected to a wireless node equipped with an IEEE 802.15.4 radio. Sensor data is transmitted to a computer that is employed for visualization. We describe practical experience encountered during the initial development of our prototype, and outline the potential battery and bandwidth conservation benefits introduced by employing popular signal processing methods. Keywords- Wireless Body Area Network; Wearable Sensors; Health Monitoring; Motion Analysis.

I.

INTRODUCTION

Wearable sensors have become viable means to monitor the health status of people diagnosed with a long-term illness or chronic condition. A small conglomerate of wireless sensor devices around a human body is commonly referred to as a wireless body area sensor network (WBASN), whereby microcontroller units (MCUs) hosting the sensor devices individually transmit digitized biological signals to an off-body gateway serving as a data relay. Off-site monitoring equipment then scrutinizes this information for subsequent assessment by medical practitioners. Nonetheless, the realization of monitoring and condition assessment solutions for home-based and ambulatory use remains highly desirable too. To this end, distinct types of sensor devices can be employed to collect vital signs data (e.g., heart rate), non-vital signs data (e.g., muscle tension), or patient daily activity (e.g., walking pace, gait, etc.) In this paper, we describe the initial development of a system using wearable sensors placed on the upper limbs of a person diagnosed with Parkinsons disease (PD). The goal of this work is twofold. First, we aim at collecting, transmitting, and analyzing PDs motion data for recognizing

the onset of dyskinesia (involuntary tremors), in order to help patients correctly time the self-administration of prescription drugs. The reason for this is that the threshold at which these drugs remain effective varies with time as the PD progresses, thus warranting a time-flexible drug therapy approach. Second, we aim at realizing a system that is scalable, easy to use, unobtrusive, and economical. To this end, system effectiveness and energy efficiency become paramount. On the one hand, we deem that the effectiveness of existing solutions [1]-[4] has been compromised by using a single sensor type. Nonetheless, an accelerometer alone does not provide information about angular motion, and a gyroscope alone does not provide information about direction of gravitational pull. A single-sensor approach thus leads to information loss that may impact the correct assessment of the patients current condition. In addition, although a dual sensor approach has been considered in [5], the importance of realizing a power-aware system that improves energy efficiency has been entirely overlooked in [1]- [5]. The main contributions of our paper are: (a) a first-hand description of practical aspects of our ongoing prototype development, and (2) a preliminary effectiveness assessment of employing the fast Fourier transform (FFT) and principal component analysis (PCA) as plausible means to gauge limb motion trends effectively, and help reduce the amount of transmitted data by the WBASN devices in order to conserve battery power through raw sensor data compression. In Section 2, we provide a brief description of our systems architecture, and describe the most relevant implementation aspects regarding sensor hardware, software interfaces and data visualization tools. In Section 3, we discuss preliminary evaluation results of our prototype. In Section 4, we discuss our ongoing research. II. SYSTEM ARCHITECTURE AND IMPLEMENTATION

A. System Architecture We aim at developing a scalable sensor platform by means of a modular hardware and software design that favours easy expandability. In other words, the WBASN shall be able to accommodate additional nodes, as well as

additional sensors on each node with minimal changes. Our initial system consists of a small group of sensor devices, a base station (BS), and a personal computer (PC). Sensor nodes comprise an MCU, a wireless transceiver, and a sensor expansion board for data acquisition, the whole of which are devised for wearing over one of the patients arm in order to monitor limb motion patterns. These nodes transmit the acquired data to the BS that is directly attached to the PC, thus forming a star network topology. All data captured by the BS is immediately forwarded to the PC for analysis and visualization. At present, all readings acquired by the sensors are transmitted as unprocessed data. Evidently, this leads to increased power consumption and high channel utilization, as discussed later in the paper. B. Sensor Hardware To streamline prototype implementation, we employed Crossbow TelosB sensor nodes, since they provide a readyto-use, onboard platform based on one of TIs MSP430 MCU family, along with the CC2420 IEEE 802.15.4 transceiver. Our previous experience indicates that this platform yields better radio range compared to the Crossbow MicaZ platform. In addition, with 10Kbytes of SRAM and 16Kbytes of program memory, the TelosB platform provides significantly more memory resources. Moreover, its physical interface is easier to work with. However, we replaced the original battery holder with a more compact one that holds a single, 3V button-cell battery. The MSP430 MCU allows additional power-saving features by offering diverse operating modes, including stand-by and deep-sleep modes. Because PD patients that follow a tightly controlled drug therapy can experience prolonged periods without dyskinesia symptoms, the MCU can instruct all peripheral devices, including the radio chip, to enter deepsleep modes of operation during such periods in order to conserve battery power by lowering active duty cycles. Then, software-triggered interrupts can re-enable the sensors to capture and analyze data periodically, as needed. To capture six degree of freedom motion information, each sensor expansion board incorporates a three-axis, digital accelerometer (Analog Devices ADXL345), and a threeaxis, digital gyroscope (Invensense ITG-3200). Hence, a data analysis subsystem can be used to correlate information acquired by both sensor types in order to implement an inertial position and limb motion pattern estimation scheme for discriminating normal movements from abnormal ones caused by Dyskinesia. Digital sensors were selected over traditional analog ones because they enable improved sensor node modularity and system scalability. Due to size and internal hardware architecture limitations, contemporary MCUs can only afford a limited number of Analog-toDigital inputs for external interfacing with diverse sensors. However, most digital sensors are easily interfaced with by employing an inter-integrated circuits (I2C) bus, which requires only two wires, a clock signal and a data channel for each individual sensor to communicate with the MCU, regardless of the total number of sensors connected, as they are all assigned unique addresses. Because sensors are daisychained together, wiring is simplified significantly.

Consequently, the future addition of sensors does not require additional I/O ports, making the system easy to scale. C. Embedded Software Interface The operation of a TelosB node is determined by user programs written in the NesC language and linked against TinyOS libraries [6] . TinyOS handles concurrency by using a split-phase approach, and provides standardized access to sensor node features and resources, such as the wireless radio and on-board data storage. The software written for our custom sensor nodes samples the digital sensors at 20 Hz. The accelerometer and gyroscope both store individual measurements for each of the three axes, which are read by accessing the sensors respective data registers, and stored in a memory buffer. As per our preceding description, digital sensor data are read by sending the corresponding commands to the sensors though the I2C channel. Our program ensures that hardware resources (i.e., sensors and the radio) are intermittently switched on and off in order to free the I2C channel to avoid access problems and to promote energy efficiency when unused. After a sensor nodes MCU obtains the respective data samples for each of the 3 axis in a sensor, it forwards them to the transceiver for immediate transmission. A single TinyOS packet contains data from the sensor measurements, header information, as well as source and sensor ID numbers that the base station employs for tracing the data origin. In addition to this, each packet also contains the reading of the MCUs internal voltage sensor, which is obtained by using TinyOS Msp430InternalVoltageC() function. All data transmissions are carried out using the TinyOS Active Message interfaces in AMSenderC. On the BS side, data received through the transceiver is forwarded to the host PC using the built-in serial port interface through a USB adapter. At present, we employ the beacon-less operation mode of the IEEE 802.15.4 standard in the sensor node radio. D. Sensor Data Visualization We developed Matlab-based algorithms and graphical user interface (GUI) for the host PC to unpack and analyze incoming sensor data received from the BS. Data is parsed, interpreted, translated to the corresponding units (m/s2 for accelerometer data, and deg/s for gyroscope data), and stored inside a cell array, grouped by node and sensor IDs. These data are then plotted on a time axis to display sensors motion readings in (near) real-time, which correspond to the node and sensor ID selected by the user. In addition, frequency-domain plots showing the outcome of FFT calculations are also provided for each of the X, Y and Z-axis. The reason for incorporating this functionality in the GUI component of the system is to explore the usefulness of identifying the presence of higher-frequency components of the incoming sensor data. However, while this can be regarded as a simple method for identifying the onset of tremors in PD patients, it serves mainly as a visual aid that helps us understand the practical traits of working with this type of motion sensors, and the overall system behaviour as we continue with our prototype improvement efforts.

(a) Accelerometer data


Time (seconds) 0 1 2 3 4 5 6 0

(b) Gyroscope data


Time (seconds) 1 2 3 4 5 6

(a) Accelerometer (reference)

(b) Gyroscope (reference)

30 20

1000 500 degrees/s 0 -500

g (m/s2)

10 0 -10 -20 -30

-1000

(c) Accelerometer (tremor onset) (d) Gyroscope FFT

(d) Gyroscope (tremor onset)

(c) Accelerometer FFT


5 4 Magnitude 3 2 1 0 0.3 2.0 3.8 5.5 7.2 8.9 10.6 12.3 14.1 15.8 17.5 19.2 Magnitude 350 300 250 200 150 100 50 0

0.2 2.0 3.9 5.8 7.7 9.5 11.4 13.3 15.2 17.0 18.9

(e) Accelerometer (dyskinesia)

(f) Gyroscope (dyskinesia)

Frequency (Hz)

Frequency (Hz)

Fig. 1. Sample plots of: time-domain (a) accelerometer and (b) gyroscope data and frequency-domain FFT analysis performed on the (c) accelerometer and (d) gyroscope, all pertaining to readings on the Y-axis during emulated dyskinesia tremors.

III.

PRELIMINARY EXPERIMENTATION RESULTS


Fig. 2. Results of applying PCA to accelerometer/gyroscope sensor outputs for: (a)-(b) normal arm motion, (c)-(d) tremor onset, and (e)-(f) full emulated dyskinesia.

As mentioned previously, we intend to produce a system whereby limb motion pattern and relative position can be estimated from the readings individually obtained by the inertial sensors. It follows that these results are timedependent. Therefore, their actual duty cycle depends on the current effect of the corresponding prescription drug on the patient. Fig. 1(a) depicts a sample reading at 20 Hz by one of our sensor nodes accelerometer over the Y-axis during emulated dyskinesia tremors (sensor node placement and arm movement are arbitrary). Similarly, Fig. 1(b) shows the gyroscopes readings during the same time period and axis. These time-domain plots clearly illustrate how large the variations in these readings can be. However, we reckon that a frequency-domain analysis of these sensors readings becomes an appealing choice, given that the distinctive tremors caused by dyskinesia could be readily recognized. This is shown in Fig. 1(c) and (d), where FFT analysis on the same a data set is performed (plots of additional sensors/axis are omitted due to space limitations). From here, it follows that, by collecting and combining results obtained through strategically-placed accelerometer/gyroscope sensor pairs, we are in a better position to characterize motion patterns that correspond to normal activity, and detect their transition into abnormal ones that can indicate the onset of a dyskinesia lapse. An important caveat here is that some normal arm movements could be mistaken for abnormal ones, especially for the cases when the PD patient is engaged in physical tasks that involve repetitive arm motion patterns (e.g., chopping vegetables at a kitchen table). This is an indication that the output of a FFT analysis alone can be difficult to

interpret by a motion pattern estimation scheme, and might yield incorrect assessments. Given the previous reasoning and the intrinsic difficulty in implementing, testing and debugging programs for embedded devices, determining an efficient and effective signal processing scheme that serves our purpose becomes crucial. We thus turn our attention to employing PCA as a complementary assessment tool to the FFT scheme. PCA provides a plausible solution to reducing the relatively large amount of sensor data that need to be collected and transmitted by transforming and reducing the original data set into a simplified, more compact one. Whereas FFT analysis converts one data type into another and can provide initial indication of abnormal limb motion, PCA implements a linear combination of the input data set, whose outcome represents trends that correspond to the variances of correlated variables. In our case, it is evident that the motion readings captured by separate pairs of inertial sensors are correlated if they are worn on the same limb. Thus, by inspecting the changes in the motion variances computed through PCA, a pattern analysis system could be in better position to accurately determine the onset of dyskinesia. Fig. 2 illustrates this through 6 different plots corresponding to the outcomes of applying PCA to accelerometer data ((a), (c), (e)), and gyroscope data ((b), (d), (f)) obtained from a separate, extended experiment. In particular, Fig. 2 (a) and (b) show the tight clustering of

TABLE I.

VARIANCES FOUND BY THE PRINCIPAL COMPONENT ANALYSIS ON THE SENSOR DATA


Principal Component
1st

Variances
Reference
0.16 0.05 0.02 71.18 23.26 5.88

Sensor type

Tremor onset
22.32 1.00 0.07 2759.39 63.29 17.51

Dyskinesia
48.31 9.27 1.34 4036.98 517.23 256.53

Accelerometer

nd rd

1st

Gyroscope

2nd 3
rd

PCAs values converted into the new coordinate system and the directions of all 3 principal component coefficients as computed by Matlab for a relatively motionless arm. Following this, a second set of measurements were taken during an emulated, minor arm tremor lapse. The results of this second experiment set are shown in Fig. 2 (c) and (d), which depict a clear scattering of the values pertaining to the outcome of the PCA over the accelerometer/gyroscope pair output from a single sensor node. Finally, Fig. 2 (e) and (f) illustrate the PCA results for data obtained during a full (emulated) dyskinesia lapse. The plots in Fig. 2 give evidence to the changes of the principal components directions revealed by applying PCA to the raw sensor measurements. Table I complements our preliminary results by showing the corresponding changes in the cumulative variances computed by PCA for both sensor types. It can be seen that the 1st and 2nd principal components account for more than 90% of the total variances for all three test cases. IV. DISCUSSIONS AND FURTHER WORK

network approach, such as the one presented in [7], could be fed with the outputs of the PCA components, along with other information in order to make an effective assessment. Another important advantage of performing signal processing computations on the WBASN nodes is that it enables ambulatory condition-assessment. To this end, the WBASN coordinator can be equipped with an on-board device to alert the user when it is time to take the prescription medication according to current motion pattern trends. However, the limited data processing capabilities of 8-bit MCUs that are widely-used in sensor nodes for reducing power consumption poses a challenge for this solution. To address this issue, employing fixed-point arithmetic functions seems a logical step towards minimizing processing time due to the lack of an arithmetic logical unit co-processor. Conversely, implementing floating-point operations through software could be done at the expense of increased memory use and longer processing time. An additional ongoing effort includes the incorporation of the beacon-enabled operation mode of the IEEE 802.15.4 radio in an attempt to further reduce power consumption at the medium access layer of the communications stack. ACKNOWLEDGMENT This project was supported by the National Sciences and Engineering Research Council of Canada under grant 365208-08. The authors thank Dr. Martin McKeown, MD, for his inputs. Ashwin Ram Sureshs work at UBC was supported by a MITACS Globalink internship. REFERENCES
[1] S. Patel, K. Lorincz, R. Hughes, N. Huggins, J. Growdon, D. Standaert, M. Akay, J. Dy, M. Welsh, and P. Bonato, Monitoring Motor Fluctuations in Patients with Parkinsons Disease Using Wearable Sensors, IEEE Transactions on Information Technology in Biomedicine, vol. 13, no. 6, pp. 864-873, November 2009. A. Salarian, H. Russmann, F. J. G. Vingerhoets, P. R. Burkhard, Y. Blanc, C. Dehollain, K. Amininan, An Ambulatory System to Quantify Bradykinesia and Tremor in Parkinson's Disease, in Proceedings of the 4th International IEEE Conference on Information Technology Applications in Biomedicine, Birmingham, UK, 24-26 April 2003. R. LeMoyne, C. Coroian, and T. Mastroianni, Quantification of Parkinsons Disease Characteristics Using Wireless Accelerometers, in Proceedings of the International Conference on Complex Medical Engineering (ICME), Tempe, AZ, USA, 9-11 April 2009. M. Pansera, J. J. Estrada, L. Pastor, J. Cancela, R. Greenlaw, and M. T. Arredondo, Multi-parametric System for the Continuous Assessment and Monitoring of Motor Status in Parkinson's Disease: An Entropy-Based Gait Comparison, in Proceedings of the 31st Annual International Conference of the IEEE EMBS, pp. 142-1245, Minneapolis, Minnesota, USA, September 2-6, 2009. S. T. Moore, H. G. MacDougall, J. M. Gracies, H. S. Cohen, W. G. Ondo, Long-Term Monitoring of Gait in Parkinsons Disease, Gait and Posture, vol. 26, pp. 200-207, Elsevier, 2007. TinyOS for Deeply-Embedded Sensor Devices. Available: http://www.tinyos.net. N. L. W. Keijsers, M. W. I. M. Horstink, and S. C. A. M. Gielen, Automatic Assessment of Levodopa-Induced Dyskinesias in Daily Life by Neural Networks, Movement Disorders, vol. 18, no. 1, pp. 7080, 2003.

Preliminary results obtained through our initial experimentations reveal the importance of employing a suitable signal processing scheme for the assessment of motion patterns in PD patients, and the potential to introduce significant bandwidth and battery conservation savings. Case in point: whereas a simple off-site data analysis approach like the one presented here requires the straightforward transmission of raw data through a bandwidth-constrained, wireless channel, this solution places a heavy burden on the button-cell batteries. In fact, we observed these batteries to last for only 2 hours of continuous operation. Conversely, by implementing the signal processing algorithms in the sensor nodes MCU, the transceivers can reduce their active duty cycle significantly. For instance, fitting the accelerometer and gyroscope readings (of all three axes), plus additional information in a single packet, each sensor node incurs 320 wireless transmissions during the course of a 16-second sample reading period @ 20 Hz. On the other hand, if PCA analysis is performed by the sensor nodes MCU immediately after a fixed number of measurements are made, then the values of all 3 principal components can be forwarded in one packet to the WBASN coordinator for subsequent combination with those obtained by other sensors in order to produce a final assessment. For instance, a neural

[2]

[3]

[4]

[5]

[6] [7]

You might also like