IRJET- IoT Based Home Automation And Health Monitoring System for Physically ...
Closed Loop DBS
1. 1
Abstract
Essential Tremor (ET) is the most common neurological movement
disorder in this country (Louis et al 1998). According to the
International Essential Tremor Foundation, it affects approximately
ten million Americans. Most patients with ET do not have tremor at
rest, but only during volitional movements (i.e., kinetic tremor).
Therapeutic deep brain stimulation (DBS), has been found to be
successful at treating ET, as well as other movement disorders such
as Parkinson's disease (Limousin et al 1999). DBS involves
implantation of electrodes into deep brain structures that are thought
to cause symptoms. Electric pulses are then delivered to these
structures to suppress pathological activity. Once stimulation
parameters are established, DBS is delivered in a continuous
manner, regardless of the status of the patient, or the severity of
his/her symptoms. In cases where the patients’ symptoms are
intermittent, continual stimulation of the brain releases unnecessary
electrical current into patients’ brains causing undesirable side
effects and unwanted tolerance build-up towards stimulation. In
addition, continuous stimulation unnecessarily depletes the batteries
of the implanted neurostimulator (INS), which in turn need to be
replaced through invasive surgery every 3 to 5 years, which is a
critical problem with the current system. Experiments with wearable
motion sensors that provide accurate data about the patient’s
movement have been used in attempt to control the DBS system in
order to mitigate these adverse effects with the patient. By controlling
the DBS system via the wearable motion sensors, stimulation only
occurs when tremor suppression is necessary (i.e. when the patient’s
upper extremities are in motion).
I. INTRODUCTION
The issues associated with continuous DBS may be
mitigated by delivering stimulation only when patients are
experiencing symptoms. Therefore, there is a dire need for
closed-loop DBS systems that can provide stimulation on
demand. The team proposes a closed-loop neurostimulation
system that automatically detects neurological symptoms
(tremors) in patients via wearable sensors, and implements
therapeutic stimulation in response to those symptoms, while
not stimulating during inactive periods. The system is
composed of a series of integrated wearable sensors that
collect user data and process it for real-time biofeedback to aid
patients with ET, in their everyday functions. The system
developed here can be easily extended to other disorders with
intermittent symptoms, such as freezing of gait in Parkinson’s
disease. In particular, the team proposes to integrate
clinical/commercial wearable sensors into the Medtronic Nexus-D
system, which allows a computer or microcontroller to send
stimulation initiation and termination signals to Medtronic clinical
DBS systems.
The Nexus-D system integration consists of three stages:
1. Integration of a clinical wireless EMG/motion sensor
system, the Delsys, Inc. Trigno System (Natick, MA) to be
used with a PC (e.g., laptop) in the loop for detecting motor
symptoms
2. Design, implementation and integration of motion
sensors with a microcontroller for PC-free operation
3. Design and implementation of a graphical user
interface to allow the physician to communicate with the
Nexus-D.
The experimental projects have been split into two
separate systems, the Home system and the Clinical system.
The systems have the ability to operate together or totally
independently.
The Home system consists of a base station and multiple
extremity motion sensor satellites. The satellite sensors
include a 9 degree-of-freedom motion sensor, MPU9250,
which includes a 3-axis accelerometer, a 3-axis gyroscope
and a 3-axis magnetometer, a small microcontroller, an
Adafruit 12 MHz Pro Trinket, and a wireless transceiver, a
Nordic Semiconductor RF24 Module (nRF24L01+) that
sends data to the base station for processing. The base
station consists of a Raspberry Pi 2 microcontroller that
interfaces with a Nexus-D neuro-stimulator and three
wireless transceivers: two for each of the satellite sensors
and one for potential communication with a PC. Power can
be fed from an electric outlet to a recharging circuitry that
feeds energy to the Lithium-Polymer (Li-Po) battery. The
user has a motion sensor attached to their body which has a
microcontroller forwarding data to the Raspberry Pi
microprocessor through a radio frequency transceiver. A
radio frequency transceiver located on the microprocessor
receives this data and processes it. Should the data indicate
Essential Tremor, the Nexus-D API is called to activate the
implanted neural stimulator and provide stimulation.
Likewise, an electric outlet feeds power to the battery
Development of a Closed Loop Deep Brain
Stimulation System using Wearable Sensors
Jackson N Cagle, Giang H Nguyen, Neel H Patel, Francy Perez, Kenan Tufekci, Dylan B Zuniga
University of Florida
2. 2
energizing the microprocessor. The Home system’s
Adafruit Motion Sensor which would process acceleration
and gyroscopic data that would be transmitted by a Nordic
RF24 transceiver to a receiver located on a Raspberry Pi
processor. This processor would interpret the information
and if movement has been detected, it would call a
stimulation method from the Nexus-D interface. A Patient
Programmer device, provided by Medtronic PLC, would
hover over the ActivaPC INS device which would deliver
stimulation. Once Nexus-D begins stimulating, ActivaPC
responds accordingly and the amount of millivolts that the
INS is inducing is reflected in an oscilloscope. Likewise,
when movement has stopped, the C++ program controlling
the Nexus-D API would cease all stimulation, making the
oscilloscope show a flat line on its screen.
For the Clinical System prototype, a graphical user
interface (GUI) that allows a physician to interact with it.
The GUI contains a series of buttons to call methods which
begin collecting data from the sensors, processing it, and
storing it. Likewise, the GUI allows the physician to deliver
stimulation and also turn it off. Several buttons are also be
present so that the physician is able to alter parameters in
the INS device such as the frequency, amplitude, and pulse
width. The the Delsys, Inc. Trigno System (Natick, MA) is
used in order to gather the motion data and the processing
algorithm is the same as mentioned for the Home sytem.
Since both the Home system and the Clinical system are
cross-compatiable, the motion sensors used for the Home
system may replace the the Delsys, Inc. Trigno System
(Natick, MA) and still have the data appear on the GUI as
well as perform stimulation when necessary.
II. MATERIALS AND METHODS (GUIDELINES FOR
MANUSCRIPT PREPARATION)
The Home system motion sensors was compiled of five
parts, an Arduino Pro Trinket 3V 12 MHz microcontroller, an
MPU-9250 9 Axis Sensor Module, a Nordic Semiconductor
nRF24L01 2.4 GHz Ultra low power transceiver, a 3.7V 500
mAh Lithium Ion Polymer Battery, and a Pro Trinket Lithium
Ion Backpack. The hardware architecture design for this is
shown in Fig 1 below.
Fig 1: Home system Motion Sensor
The battery is connected directly to the backpack, which
enables it to be recharged at will. The code for the
microcontroller was written in C++. The battery is also used to
power the other three components, the microcontroller, motion
sensor, and wireless transceiver. As can be seen in Figure 3,
the voltage output of the battery is 3.3V, but it is stepped
down to 3V in the regulation circuitry built into the
microcontroller.
The Home system’s motion sensors are able to
communicate wirelessly with the Base Station using the
transceiver on board. The Base Station consists of four
separate components, being a Raspberry Pi 2 for processing,
another Nordic Semiconductor nRF24L01 2.4 GHz Ultra low
power transceiver, a 3.7V 6600mAh Lithium Ion Battery
Pack, and an LB-Link 802.11N USB Wifi Module for
communication with the computer in the Clinical System. The
Hardware Design Architecture for the Base Station is shown
in the Fig 2 below.
Fig 2: Base Station
As with the Home system, the battery on the Base Station
powered all components, although it did not need to have its
voltage stepped down. The Base Station was hardwired to the
Nexus D via a USB cord for direct communication. Also, the
code for the Raspberry Pi 2 was written in Raspbian, a
derivative of Python. The Base Station, being a much simpler
design than the Mobile Station, did not necessitate it's own
PCB. Instead, the battery was directly attached to the bottom
of the Raspberry Pi using an epoxy glue, while the remaining
two components were hardwired to the GPIO pins. This entire
set up was then able to sit in the front pocket of the user, or
alternatively attached to an external harness.
The final aspect of hardware for this system was the Clinical
System. When the user is in the office of a clinician, and the
system can switch from Home Mode to Clinical Mode, the
Base Station is longer needed to communicate with the Mobile
Station. Instead, the Base Station will receive commands from
the PC of the clinician. Also, rather than using the sensors
from the Mobile Station, the Clinical System utilizes Delsys
Trigno Sensors, which include accelerometers and EMGs on
them.
III. RESULTS
On each extremity, the patient wears a simple bracelet
containing a motion sensor set. The bracelet collects
movement data using accelerometers, gyroscopes, and
magnetometers. Using a wireless transmitter, each bracelet
3. 3
sends motion data to the base station for processing. As
described previously, the motion data collected by each
wireless motion sensor is transmitted through the wireless
transceiver to the processing base station. The motion data
collected by the motion sensors are each sampled by a 16-bit
Analog-to-Digital Converter with the most significant bit
representing the sign. After converting to the new referenced
space, three 4-byte (single precision) acceleration data and
three 2-byte (unsigned integer) rotation data is packaged with
4-byte (unsigned long integer) time reference and 2-byte
(unsigned character with unused package space) header
information together as one single payload for transmission.
Therefore, each sensor is transmitting 2400 bytes-per-second
to the base station. However, due to hardware limitations of
the RF24 series chip, each transceiver module can only
send/receive data one channel at a time. With multiple
transceivers activating in multiple locations of the body,
wireless interferences occurs. Multiple methods are tested and
compared as described later in this report. To combat the
wireless interference problem, the Relay method is
implemented on each wireless motion sensor as shown in
Figure 1. Although only two sensors are shown in the figure,
the Relay method can be used with more than 2 sensors for
each receiving module on the base station. However, delay
and complexity increases as the number of sensors increases.
Fig 1: Relay System
In the Relay method, Motion Sensor 1 transmits its 24 byte
motion data to Motion Sensor 2 in transmission channel 1.
Motion Sensor 2 receives the data, concatenates it with
Motion Sensor 2’s motion data, converts the transmission
channel from channel 1 to channel 2, and then transmits 48
bytes of motion data to the base station. After transmission,
Motion Sensor 2 converts the transmission channel from
channel 2 back to channel 1 and wait for incoming data from
Motion Sensor 1. If one of the sensors is deactivated due to
unknown reasons, a contingency plan is executed, which starts
transmission as if there is only one sensor. The conditions for
contingency are if Motion Sensor 1’s data filled up the
transmission buffer because Motion Sensor 2 does not receive
data properly, and if Motion Sensor 2 does not receive data
from Motion Sensor 1 for more than 10 consecutive data
points.
In order to measure the actual extremity tremoring of the
patients, a baseline correction is essential to remove false
detection due to possible acceleration interferences with
automobile or other possible causes. Therefore, a motion
sensor is added to the base station for overall baseline
acceleration. Rotation information is not affected by this
information; in turn, gyroscope data is not be affected for the
baseline correction. The major problem with the base station
motion sensor is that it does not necessarily have to be in the
same coordinate system as the wireless motion sensor because
the reference is the integrated circuit orientation. The solution
to this problem is a projection algorithm based on Attitude and
Heading Reference System (AHRS). AHRS, as described in
Fig 2, is traditionally used for aircraft control. In current
design, AHRS is used as the standard coordinate for all motion
measurement in this application. In order to convert the
coordinate of each wireless motion sensor to the standard
coordinate system, the Directional Cosine Matrix (DCM) is
applied to the system during each sample calculation. All
AHRS computation is using magnetometer for Yaw
computation instead of traditional gyroscopic computation
since the magnetometer's reference is more accurate.
Fig 2: Altitude and Heading Reference (AHRS)
Fig 3 shows the initial data collected from the motion sensors
as well as the frequency distribution of the data.
Fig 3: Walking vs. Tremoring Raw Data
Fig 4 shows the base line corrected data for gravity that was
applied to the data that was collected and shown in Fig 3. Both
the data as well as the frequency distribution is shown.
4. 4
Fig 4: Walking vs Tremoring Base Line Corrected Data
A comparison of the raw data to the corrected data shows that
the gravity effect was corrected while maintaining the same
frequency components.
Among many possible methods of detecting tremors,
Support Vector Machine (SVM) stands out for its both high
accuracy and easy implementation. SVM uses information
from baseline corrected acceleration data and rotation data to
predict the current condition of the user. Five essential
features are required for SVM, including 1) acceleration
frequency at 8 to 12 Hz, 2) rotation frequency at 8 to 12Hz, 3)
variance of acceleration, also known as the power, 4) variance
of rotation, 5) classification results of the previous sample.
These features are computed for every 250ms window, which
is the minimum window to compute the frequency at 4Hz
resolution in Fourier Transform without using the overlapping
or the windowed method. It is commonly recognized that the
overlapping or windowed method requires at least 2 to 3
magnitudes of longer computation time than normal Fourier
Transform; thus, only traditional discrete Fourier Transform is
used. Due to the small window size, the classifier might
misclassify sudden movement as tremoring. Therefore, the
previous window’s classification results is used as a correction
for possible false detection. A 10-fold cross-validation result
from over 20 minutes of training data indicates an average of
97.8% accuracy as shown in Fig 5.
Fig 5: Support Vector Machine
IV. DISCUSSION
The results from the Multiple Sensor Communication
testing indicate the relay method for Radio-Frequency
communication is the best method for multi-sensor data
collection. The direct transmission of two sensors causes too
much interference for each other, and the relay method
counters this problem by using different frequency bands for
Sensor-to-Sensor and Sensor-to-Base communication.
Although the receiver scanning method also uses different
frequency bands, the alteration of frequency leads to a loss of
package when the transmitter stops transmission after the
package is sent. There is no error-checking feedback in the
Nordic Radio-Frequency chip, so the wireless motion sensor
stops its transmission after first broadcasting and the receiver
fails to receive data during frequency switching.
The SVM model was introduced, implemented, and proved
that there is about 98% accuracy of classifying the data given
the training set. As of right now, SVM proves to be the most
efficient way of classifying the data while still maintains a
high level of accuracy.
Since this system shows that it can increase the efficiency
and effectiveness of the current open loop DBS treatment
method, steps can now be taken to consolidate the hardware
and software of this system into the next generation of Nexus-
D and INS technology. If this system can be implemented in
such a way that is not cumbersome to users, then the quality of
life of individuals with Essential Tremor, and potentially those
with other disorders in which this system can be applied to,
will significantly increase.
V. OTHER RECOMMENDATIONS
Another classification method such as 20 layer Neural
Network can be implemented to have more accurate
classification. The system should be integrated at varied
frequencies to determine effectiveness with other movement
disorders, such as Parkinson's disease.
REFERENCES
[1] Louis ED, Ottman R and Hauser WA “How common is the most
common adult movement disorder?: Estimates of prevalence of essential
tremor throughout the world” in Movement Disorders 1998; 13:5-10
[2] Limousin P, Pollak P, Van Blercom N et al. (1999). Thalamic,
subthalamic nucleus and internal pallidum stimulation in Parkinson's
disease. J Neurol 246: 42-45