23-08-2012, 03:23 PM
Patient Fall Detection using Support Vector Machines
patient FallDetection.pdf (Size: 402.68 KB / Downloads: 39)
Abstract
This paper presents a novel implementation of a patient fall detection
system that may be used for patient activity recognition and emergency
treatment. Sensors equipped with accelerometers are attached on the body of
the patients and transmit patient movement data wirelessly to the monitoring
unit. The methodology of support Vector Machines is used for precise
classification of the acquired data and determination of a fall emergency event.
Then a context-aware server transmits video from patient site properly coded
according to both patient and network status. Evaluation results indicate the
high accuracy of the classification method and the effectiveness of the
proposed implementation.
1 Introduction
The telemonitoring of human physiological data, in both normal and abnormal
situations of activity, is interesting for the purpose of emergency event detection or
long term data-storage for later diagnosis or for the purpose of medical exploration.
In the case of elderly people living on their own, there is a particular need for
monitoring their behavior. The first goal of this surveillance is the detection of major
incidents such as a fall, or a long period of inactivity in a part of their area. The early
detection of fall is an important step to alert and protect the subject, so that serious
injury can be avoided. Fall detection is an important part of human body movement
analysis; it is considered as an area of increasing importance and interest to
practitioners, researchers, and health industry and most importantly, it is vital for
indication of emergency cases. Accelerometers have been proposed as a practical,
inexpensive and reliable method for monitoring ambulatory motion in elderly
2 Charalampos Doukas1, Ilias Maglogiannis1, Filippos Tragkas2, Dimitris Liapis2,
Gregory Yovanof2
subjects for the detection and prediction of falls. Robust classification of motion and
postures from accelerometer data enable the development of more reliable methods
for monitoring long term change in physiological indicators such as parameters of
gait, balance, energy expenditure and general well-being.
This paper presents a patient fall detection platform based on accelerometer data.
Body sensors collect the movement data and transmit them wirelessly to the
monitoring unit. Appropriate data classification using Support Vector Machines [19],
can classify the recorded movement into three categories; fall, walk and run. Then
the deployment of additional context awareness based on the previous activity
detection may enable the proper coding and transmission of video images from the
patient to remote monitoring units (i.e alarm triggering and high quality video
transmission). The rest of the paper is organized as follows; Section 2 discusses
related work in the context of patient activity and fall detection. Section 3 describes
the acquisition of the patient movement data using sensors, whereas Section 4
presents that data classification using Support Vector Machines. The whole system
architecture is described in Section 5 and Section 6 presents the evaluation results.
Finally, Section 7 concludes the paper.
2 Related Work
Although the concept of patient activity recognition with focus on fall detection is
relatively new, there exists related research work, that may be retrieved from the
literature ([1]-[16]). Information regarding the patient movement and activity is
frequently acquired through visual tracking of the patient’s position. In [6] and [15]
overhead tracking through cameras provides the movement trajectory of the patient
and gives information about user activity on predetermined monitored areas. Unusual
inactivity (e.g., continuous tracking of the patient on the floor) is interpreted as a fall.
Similarly, in [10] omni-camera images are used in order to determine the horizontal
placement of the patient’s silhouettes on the floor (case of fall). Success rate for fall
detection is declared at 81% for the latter work. Head tracking is used in [13] in
order to follow patient’s movement trajectory with a success rate of fall detection at
66.67%. The aforementioned methods that detect falls based on visual information of
the user require capturing equipment and thus are limited to indoor environment
usage. In addition, some of the methods require also the a-priori knowledge of the
area structure (e.g., obstacles, definition of floor, etc.), or user information (e.g.,
height in [10]). A different approach for collecting patient activity information is the
use of sensors that integrate devices like accelerometers, gyroscopes and contact
sensors. The decrease of sensors size and weight, in conjunction with the
introduction of embedded wireless transceivers allows their pervasive placement on
patients and the transmission of the collected movement information to monitoring
units wirelessly. The latter approach is less depended on the patient and
environmental information and can be used for a variety of applications for user
activity recognition ([1], [3], [9]). Regarding fall detection, authors in [2], [7], [8],
[12] use accelerometers, gyroscopes and tilt sensors for movement tracking.
Patient Fall Detection using Support Vector Machines 3
Collected data from the accelerometers (i.e., usually rotation angle or acceleration in
the X, Y and Z axis) is used in order to verify the placement of the patient and time
occupation in rooms and detect abrupt movement that could be associated with fall.
Detection is performed using predefined thresholds [1], [3], [4], [7] and association
between current position, movement and acceleration [2], [8], [12]. Finally, area
sensors have been used in order to track and analyze patient movement; authors in
[11] describe a vibration-based detector that can detect falls based on the vibration
caused on the floor. In [5] infrared sensors are used that provide thermal information
regarding the patient’s location and movement. The latter approaches do not require
from the user to wear or carry sensor devices, however they demand more expensive
equipment to be installed on the surrounding environment.
Most of the related work based on accelerometers for fall detection, focuses on the
elderly and may not be used for general classification of the patient movement
activity or usage in younger ages (e.g. interpretation of running has not been assessed
against falling). In addition, detection is usually performed through predefined
thresholds and thus results can be depended to the movement patterns of the users.
The presented system is using a state of the art classification methodology, the
Support Vector Machines, for data classification and fall detection. The proposed
system may be used for a variety of patient activity recognition since it can
successfully distinguish movement between run and fall. In addition it is not biased
by the movement pattern or the physiology of a specific patient (i.e. it can perform
successfully with movement data from different individuals) and it does not apply
restrictions to the user’s environment (e.g., it can be used in outdoor environments as
well). A context-aware framework deployed within the system enables the proper
transmission of video images from the patient in case of emergency events,
optimizing the whole telemonitoring procedure.
3 Patient Movement Data Acquisition
This section provides information on the acquisition and pre-processing of the
patient movement data. The MC13192 [2] sensor has been used in our system. The
latter contains a 2.4 GHz wireless data transceiver RF reference design with printed
circuit antenna, which provides all hardware required for a complete wireless node
using IEEE 802.15.4 (ZigBee) [17] packet structure. It includes an RS232 port for
interface with a personal computer, background debug module for in-circuit
hardware debug, four switches and LEDs for control and monitoring, a low-power,
low-voltage MCU (MicroController Unit) with 60KB of on-chip Flash which allows
the user flexibility in establishing wireless data networks and two 3D Accelerometers
for X, Y and Z axis. Fig. 1 shows the SARD ZigBee node [18]. ZigBee has been
chosen as communication technology for a number of reasons:
• Low cost and very low power consumption: its data rate is at 250 kbps and
its power consumption is 30mA in Transmit mode and only 3μA in StandBy
mode respectively. A ZigBee node can thus have a very long battery life (2-
3 years with a AA cell).