0
System for Remote Multichannel Real-Time
Monitoring of Mouse ECG via the Internet
Matthew Blake Oefinger
B.S., Electrical Engineering
B.S., Computer Science
B.S., Mathematics
Southern Methodist University
SUBMITTED AUGUST 2003 TO THE DEPARTMENT OF ELECTRICAL ENGINEERING
AND COMPUTER SCIENCE INPARTIAL FULFILLMENT OF THE DEGREE OF
MASTER OF SCIENCE AT THE MASSACHUSETTS INSTITUTE OF TECHNOLOGY
Copyright Massachusetts Institute of Technology. All rights reserved.
Author Signature:
of Electrical Engineering and Computer Science
Department
Department of Electrical Engineering and Computer Science
August 25, 2003
Certified by:
Dr. Roger Mark, M.D., Ph.D.
Distinguished Professor in Harvard-MIT Health Sciences and Technology
Professor of Electrical Engineering
Department of Electrical Engirearing and Computer Science
Accepted by:
Dr. Arthur C. Smith, Ph.D.
Professor Emeritus of Electrical Engineering
Chairman, Committee of Graduate Studies
MASSACHUSEUTS INS.TiTUTE
OF TECHNOLOGY
OCT 1 5 2003
ARCHIVES
LIBRARIES
System for Remote Multichannel Real-Time Monitoring of
Mouse ECG via the Internet
Matthew Blake Oefinger
Submitted August 2003 to the Department of Electrical Engineering and Computer Science
inpartial fulfillment of the requirements for the degree of
Master of Science in Electrical Engineering
ABSTRACT
A hardware/software system was developed to allow real-time monitoring of multiple
physiological signals simultaneously via the Internet. The hardware is specifically designed for
measuring ECG signals from mice, while the software system is agnostic to the underlying data
source. The software utilizes a client-server model and multicasting network technology to
stream real-time data from a server to clients. The server software includes an intuitive
graphical user interface (GUI) that allows lab technicians to start and stop data collection as
well as a back-end SQL database for record archiving and management.
The system is intended to facilitate real-time signal processing development as well as
collaborative research of archived physiological data among geographically disjoint groups.
The server is therefore equipped with a digital signal processing (DSP) board for real-time data
analysis and password-protected access to physiological data archives.
Thesis Supervisor: Roger Mark
Title: Distinguished Professor in Harvard-MIT Health Sciences and Technology
Professor of Electrical Engineering
1
Introduction
Years of cumulative research have led to modern medical dogma stating that high blood
levels of low-density lipoprotein (LDL) coupled with low levels of high-density lipoprotein
(HDL) impel atherosclerosis and subsequent cardiac maladies (1). Factors contributing to
lipoprotein imbalances include genetic predisposition, poor diet, and inadequate exercise.
In this study, part of a collaborative effort between MIT Health Sciences and Technology
Division and the Department of Biology, I designed and built a data acquisition system
and a client/server software architecture for remote monitoring of the electrocardiogram
(ECG) signal of mice genetically engineered with LDL and HDL metabolic deficiencies (and
consequently atherosclerosis and diseased hearts). The system consists of three components,
each of which will be described in detail in this paper:
1. 16-channel ECG hardware
2. server software for data acquisition, archiving, and parallel broadcasting of 16-channel
ECG signals
3. client software for remote real-time monitoring (visualization) of the ECG signals
Ultimately Professor Monty Krieger in the Department of Biology will use this system to
research trends of cardiac disease in mice by correlating electrophysiology with histological
pathology. Specifically, the real-time monitor will be used as a tool to alert biologists to
the presence of cardiac events of interest (arrhythmia, ST elevation/depression, etc.) at
which point they will sacrifice the subject mouse to obtain cardiac tissue samples, thereby
correlating ECG events with histological data. They will eventually also try therapeutic
interventions, using ECG monitoring as a basis for understanding the effects of the intervention. Such advanced correlations will only be possible with accurate real-time ECG signal
processing algorithms, which serve as the foundation for my future studies.
2
2.1
Background and Motivation
apoE and SR-BI: The Double Knockout Model
Since 1900, cardiovascular disease has been the number one killer in the United States every
year but 1918, and atherosclerosis is the major cause of heart disease and stroke (1,2).
Risk for atherosclerosis is proportional to LDL cholesterol levels in the body, while risk is
inversely proportional to HDL cholesterol. Michael Brown, Joseph Goldstein and colleagues
established an understanding of receptor-mediated metabolism of LDL beginning in the
1970s (3,4), while only as recently as 1996 did researchers unveil the existence of a HDL
receptor and begin to comprehend fully the associated pathways of HDL metabolism (1,3).
For decades researchers have sought to breed murine subjects with atherosclerosis in an attempt to model the human disease. Original mouse models, known as "single-knockout"
(KO) mice, were genetically engineered with a homozygous null genotype encoding the segment of DNA that would normally encode apolipoprotein E (apoE). Because apoE is a critical
enzyme in the metabolism of LDL, single-knockout mice are characterized by severely elevated LDL levels. A single-knockout mouse, however, does not typically exhibit many of the
defining features of human coronary heart disease (CHD), namely spontaneous myocardial
infarction, severe cardiac dysfunction, and premature death (4).
With increased understanding of the biochemistry of HDL and its associated role in physiological systems came emphasis of the joint importance of LDL and HDL as contributors
to CHD. Researchers at MIT and Harvard Medical School developed a "double-knockout"
(dKO) mouse, engineered to lack DNA encoding apoE and SR-BI, a critical molecule for
cellular absorption of HDL (4). The double-knockout mouse exhibits symptoms more nearly
modeling CHD in humans, namely lipid-rich coronary artery occlusions (fig. li), multiple
myocardial infarctions, an enlarged heart (fig. lii), reduced ejection fraction and contractility
(figs. liii,2), ECG abnormalities (figs. 3i,3ii,3iii), and premature death (fig. 4).
A
Angiograms
wild type
dKO
B
100
6
75
T
0
apoE KO
dKO
50
25
0.
apoE dKO
KO
Figure 1:
(i) Angiogram comparison of wild type (left) vs. apoE/SR-BI dKO (right) mice. The arrows indicate left anterior descending
(LAD)coronary artery. The DKO mouse shows multiple severe stenoses in the coronary arteries branching from the LAD.(ii) B and C, Posterior
views of perfused hearts from apoE KO (6 weeks old, B) and dKO (5 weeks old, C) mice. D, lateral view of right side of heart from a different
5-week old dKO mouse. Black and white arrows represent left and white atria, respectively; arrowheads, myocardial lesions. Original magnification:
7.5X. (iii) MRI-based ejection fractions from control apoE KO and dKO mice. [Reprinted with permission (6)]
2.2
Electrophysiology: ECG and Cardiac Arrhythmia
The electrocardiogram is an invaluable tool in understanding cardiac activity and diagnosing deviations from healthy behavior. A quintessential human ECG shows that a single
HemdynarnI- Analyseaof Hart Function InControl and dKO Mbe
Aortic Systolic Aortic D iastolic
Pressure,
Pressure,
LVEDP,
LVSP,
+dP/dt,
-dPidt,
Genotype
mm Hg
mm Hg
mm Hg
mm Hgs
mm Hg
mm Hg
Wild Type
86+-17 (6)
61+-17 (6)
5.7+-0.9 (6) 88+-17 (6) 3800+-900 (6) -3400+-800 (6)
SR-BI KO
79+-20 (5)
55+-22 (5)
7+-4.4 (3)
73+-11(S)
:200+-900 (3) -2900+-600 (3)
apoE KO
77+-9 (5)
52+-12 (5)
6.1+-1.7 (5)
82+-7 (5)
500+-500 (5) -3300+-500 (5)
524-+-77 (5)
dKO
54+-5 (6)
39+-S (B)
9.6+-2.2 (6) 40+-19 (6)
1100+-500 (6) -1100+-400 (6)
390+-39 (6)
Heat Rate, bpm
509+-71 (6)
565+-122 (5)
Figure 2:
LVEDP indicates LV end diastolic pressure; LVSP, LV systolic pressure. Values represent mean ±1 SD. Numbers of animal per
group are indicated in parentheses. Reprinted with permission (6) [details of HR measurement method not provided in (6).]
heart beat consists of P,Q,R,S and T components, where the P-wave corresponds to atrial
depolarization, the QRS complex is indicative of ventricular depolarization, and the ST segment represents ventricular repolarization. In a mouse ECG, by contrast, the T-wave is not
discernable (fig. 5).
In many cases one may discern the nature of a cardiac lesion based on its ECG signature.
While a thorough understanding of all the intricacies of ECG interpretation is reserved for
cardiologists and others who devote extensive effort to their art, one may glean invaluable
information by focusing on the ECG signature of cardiac rhythm. An abridged description
of the mechanism of cardiac rhythm, adapted from Dubin's classic textbook (5), follows.
Electrical activity of the heart (fig. 6) begins at the SA node, the heart's natural pacemaker,
that generates a sequence of action potentials controlling the rate and rhythm of the heart.
The action potential spreads from the SA node to the atrial myocardium, the depolarization
of which creates the P-wave on an ECG recording. Depolarization next spreads to the AV
node where its conduction velocity slows down dramatically, permitting atrial contraction
to complete before ventricular contraction begins. The action potential next spreads rapidly
through the Bundle of His and the ventricular conduction system. Purkinje fibers distribute
the action potentials to the inner wall of the ventricles, thus initiating ventricular depolarization and a corresponding QRS complex on the ECG. Repolarization of the ventricles leads
to the T-wave. The currents associated with the spread of depolarization through the AVN
and ventricular conduction systems are too small to generate visible wavelets on the surface
ECG. As a result the period between the end of the P-wave and the beginning of the QRS
is relatively flat, and is called the isoelectric level.
The SA node typically acts as the heart's primary pacemaker, but it can, for various reasons,
fail to generate the requisite impulse to stimulate a heartbeat. Under such circumstances,
another automatic focus, with a slower inherent pacing rate, compensates as a backup sys-
(i)
Patient: mars-dko..11_1,
ID:
ST Segment Depression
(ii)
Patient: marsdko_11_1,
ID:
AV Block
T
ZOfi V*G
10
M"/VPause 2380 ms
N-N
02-Sep-2002 09:49:50
60BP
6.p
IV
(iii)
Patient: mars-dko_ 11,
ID:
JEB
10omm/rn
(No rata)
10 MmhV
....
N-N
...Pause 2040 ms
02-Sep-2002 16:51:19
65 BPMW
(iv)
Figure 3:
(i) Normal ECG (ii) ST Segment Depression (iii) A/V Block leading to 2 missed beats (iv) Missed beat followed by junctional
factor of 10, then utilizing MARS® software for automatic arrhythmia
escape beat [Above data was scanned by time-expanding mouse data by
detection and annotation.]
4
Age (weeks)
Figure 4:
Survival (%) as a function of time. The average lifespan of a healthy mouse is approximately 2 years, while 100% of the DKO
population has died by 8 weeks of age. [Reprinted with permission (6)]
v5 /
T
P
U
QS
Figure 5:
ST
segment
......
R.
PR
QT
interval interval
(left) Normal human ECG. (right) Normal mouse ECG (note the absence of a T-wave)
tem. There is a distinct hierarchy of dominance, based on a descending inherent pacing rate,
dictating which automatic focus will emerge to provide electrical stimulation. Anatomically
the dominance descends from the atria, through the A/V junction and into the ventricles.
Within a given anatomical region multiple automatic foci vie for dominance since they share
similar inherent rates.
A lapse in pacemaking from the SA node, or any automatic focus for that matter, may be
transient or permanent. In the event that the failure is transient, once the superior healthy
automatic focus regains function it overdrive-suppresses all slower foci, and the inferior
automatic focus merely generates one or several escape beats (fig. 7) at its inherent
pace. In the event that the failure is permanent, the inferior automatic focus becomes
the dominant pacemaker, generating an escape rhythm. Though ideally the hierarchy of
dominance in automatic foci is obeyed, in a diseased heart one may find that blocked electrical
conduction pathways prevent foci from receiving overdrive-suppression signals from other
foci, causing multiple automatic foci to fire nearly simultaneously, a phenomenon known as
multifocal ectopic firing. In the case that multifocal ectopic firing occurs in the atria,
atrial fibrillation can ensue. In a severely diseased heart in which conduction superior to
the ventricles is blocked, ventricular multifocal ectopic firing can lead to lethal ventricular
fibrillation. Examples of fibrillation are shown in figure 8.
Figure 6:
Simplified diagram of the heart conduction system: the SA node, "the pacemaker," is positioned on the posterior wall of the right
atrium. SA node action potentials lead to depolarization of the atria. The action potential descends from the SA node through the internodal
tract to the AV node, then the Bundle of His and subsequently the bundle branches. The action potential continues to propagate from the bundle
branches to Purkinje fibers, which causes depolarization of the ventricular myocardium.
(ii)
(i)
(iii)
Figure 7:
Examples of escape beats in human ECG: (i) atrial escape beat (could also be a junctional escape beat with atrial retro-polarization)
(ii) junctional escape beat (note the absence of a P wave (iii) ventricular escape beat (note the widened QRS complex)
Even while limiting ECG analysis to the scope of arrhythmia detection, one finds that arrhythmias can stem from an ample variety of causes, some of which are nearly indistinguishable on an ECG recording. While an interpretation of the characteristic signature of different
arrhythmias as a diagnostic tool for an underlying cause is an invaluable clinical tool, the
motivation for arrhythmia detection in the context of this study is not yet so specific. Eventually we will perform more advanced studies to correlate specific ECG characteristics with
histology studies. A necessary foundation for such a study, however, is an instrumentation
and signal acquisition/archiving system that can reliably record multiple ECG signals with
high fidelity for weeks at a time. As we were unable to find a system adequate for our
specific needs, the design of such a system provided the short-term goal upon which I based
my master's work; the challenges of more advanced signal processing and real-time ECG
analysis for correlation of electrophysiology with histology await future studies.
NormaI ECG
Atrba-l Fibrillation ECG
Ventdculcw lbilaton
(i)
(ii)
Figure 8:
examples of fibrillation in human ECG: (i) atrial fibrillation: P waves are distinctly absent, as the atrial myocardium has no
coherent polarization pattern when multiple atrial automatic foci create simultaneous stimuli (ii) ventricular fibrillation: the expected Q,R,S and
T wave segments are not discernable, as the ventricles have no coherent polarization pattern due to multifocal ectopic firing.
3
3.1
ECG Data Acquisition System
Acquisition Hardware
This section details the design architecture and process used in building 16-channel hardware
for mouse ECG signal conditioning. The final machine, shown in figure 9, is approximately
16" x 9" x 9" in size. It is cooled by two brushless DC fans (not shown) that emit minimal
signal-degrading electromagnetic radiation when compared with their AC counterparts. The
lead cables terminating at the input terminals of the analog cards are enclosed in braided
metal shielding which is grounded to reduce capacitively-coupled noise. The output of each
of the 16 analog signal conditioning cards is wired into a DSP termination card (mostly
blocked by the analog cards and therefore barely visible in figure 9). This DSP termination
card is connected to a DSP board on the data acquisition RAID server via a ribbon cable,
where the conditioned (filtered and amplified) analog signals are digitized and stored. A
signal obtained with the hardware is shown in figure 10.
The signal conditioning cards are built with sockets into which microchips are inserted. Such
modularity allows for easy adaptation of the system to other uses. Consider, for example,
the case in which the system is to be utilized for human ECG signal conditioning. Simply
removing the filters and replacing them with those appropriate for the human ECG spectrum,
with no de-soldering and re-soldering of components required, allows for an easy change. The
modularity of the hardware has the added advantage that debugging is greatly simplified.
Again due to the socket-connected design of the analog cards, a single part may be swapped
without soldering. Such ease of swapping reduced the design time substantially by allowing
faulty parts to be identified and replaced efficiently.
Figure 9:
16-channel ECG hardware implementation designed for Krieger Lab murine studies.
Figure 10:
Mouse ECG signal acquired using ECG hardware described in this section.
3.1.1
Single-Channel Architecture
The below diagrams (Figures 11, 12, 13, 14, 15) show the overall ECG instrumentation
architecture. The entire acquisition system consists of 16 such instruments operating in
parallel.
9
Figure 11:
This schematic shows a circuit for measuring a single-lead ECG signal. One electrode (INPUT C) serves as ground while the
voltage difference between the other electrodes (INPUTS A and B) serves as the signal input. The Bessel filters, which were purchased as modules,
are in fact complicated analog circuits but are abstracted as black boxes in the above schematic. As following diagrams show, this schematic is then
converted into a PCB (printed circuit board) layout, which includes all necessary hole dimensions and trace connections to allow manufacturers to
implement the board.
Figure 12:
This schematic, which comprises a single channel of the ECG instrumentation, was created using PCB1239 software. As
following diagrams show, this schematic is then converted into a PCB (printed circuit board) layout, which includes all necessary hole dimensions
and trace connections to allow manufacturers to implement the board.
10
Figure 13:
The schematic (shown in previous picture) is converted into a PCB (printed circuit board) layout diagram after all microchip
dimensions are entered. From this diagram manufacturers can fabricate a printed circuit board to realize the schematic.
Figure 14:
3.1.2
Fabricated printed circuit board of single-channel ECG hardware.
Primary Common-Mode Noise Reduction: Active Grounding Circuit
Sixty-hertz electromagnetic noise (and to a lesser extent harmonics thereof) is ubiquitous
indoors, as electrical systems in buildings utilize AC power delivered at 60 Hz (50 Hz in
European countries). Because the spectrum of mouse ECG spans 0-1kHz1 , using a notch
filter to remove 60 Hz noise will invariably remove at least some signal content. An active
ground circuit (fig. 17) is the preferred means of removing such common-mode noise.
'According to one study (9), the frequency range 0-80 Hz contains 45.48 +/- 15.32% of the spectral
power, while the range 80-1000 Hz contains 54.52 +/- 15.32% of the spectral power. That is, nearly all the
Optical Instrumentation Amp
HPF
opamp
anhliasing LPF
Figure 15:
Simplified Diagram of Hardware Setup: Mouse ECG signal amplified with optically-isolated instrumentation amp followed by a
highpass filter, second amplification stage, then a lowpass antialiasing filter. The signal is finally sampled by a Microstar@ Laboratories model
4200A A/D card (not shown).
In the model in figure 16, R3 and R4 represent the skin/electrode impedances of the + and
- (signal) electrodes, while RI and R2 represent the Thevenin equivalent input impedances
as seen at the + and - terminals, respectively, of the preamplifier. For the purposes of this
model it is fair to treat R1 and R2 as equal, since nonideal opamp characteristics are secondorder effects in the understanding of noise. It is also fair to treat Il and 12 as equal, since
the strength of electromagnetic radiation, which produces the capacitively-coupled currents
Il and 12, will not vary significantly over the small distance between electrodes on a mouse.
R3 and R4, however, are very unlikely to be identical, since skin/electrode impedances will
vary due to subject movement, initial placement, anatomical differences in the respective
implantation areas of the electrodes, etc. Therefore since R1/R3 is different from R2/R4,
the common-mode capacitively-coupled current produces a differential voltage across the
input terminals of the preamplifier. This differential voltage produces a noise signal that
is propagated through to the output, even for an ideal opamp with infinite common-mode
rejection.
An active grounding circuit, modified from a description by Huhta and Webster (17), is
shown in figure 17. The circuit works by taking the average (common mode) of the voltages
at the two input terminals of the preamplification stage, amplifying and inverting the signal,
and feeding that signal back as the ground, or reference voltage, for the circuit. The circuit
does not remove differential signal content but mitigates common-mode noise, that is signal
that is simultaneously present on both electrodes. With an active grounding circuit included
in the architecture, a high-quality signal results, as figure 10 shows.
3.1.3
Secondary Common-Mode Noise Reduction: Increasing Preamplifier CMRR
with an Instrumentation Amplifier
Active grounding is described by Huhta and Webster (10) as the optimal means of removing
common-mode noise, of which 60 Hz noise is most predominant. Recall that common-mode
60 Hz noise is present independent of the common-mode rejection capabilities of the amplifier.
signal power resides in the 0-1000 Hz range, motivating the chosen passband of 0.1 Hz - 1000 Hz.
>3 R1
3650KG
R33
+
6
2/
2R
4R2
PREAMPLIFIER
Figure 16:
A model explaining why common-mode 60 Hz noise contaminates ECG data: skin/electrode impedance deifferences between
signal electrodes. R3 and R4 represent the skin/electrode impedances of the + and - electrodes as shown in figure 15, while R1 and R2 represent
the Thevenin equivalent input impedances as seen at the + and - terminals of the preamplifier, respectively.
As a second-order consideration, however, it is important to understand the role that the
common-mode rejection ratio (CMRR) of an amplifier plays in reducing common-mode noise.
Even if, in figure 16, R1/R3 were exactly equal to R2/R4, a small amount of common-mode
noise would, in reality, still be present at the output. Such is the case because opamps do
not have infinite CMRR, as the model in figure 16 assumed.
An instrumentation amplifier increases the CMRR of the system. As shown in figure 18, an
instrumentation amplifier consists of three opamps. Consider first a common-mode voltage
input to the first stage of the instrumentation amplifier circuit; that is, Va = Vb in figure 18.
If Va = Vb, then VI = V2 also, by the virtual-ground property of ideal opamps. The virtualground property states that, since for an ideal opamp the open-loop gain is infinite, then
in a negative feedback configuration the positive and negative terminals of an opamp must
be equal in voltage in steady-state. Since V1 is equal to V2, then no current flows through
R1. Knowing that the input impedance at the terminals of an ideal opamp is infinite, by
Kirchoff's current law the current through R2 as well as R3 is zero. Therefore V1 = V2 =
Vouta = Voutb. The common-mode output, calculated as (Vouta + Voutb)/2, is therefore
equal to the common-mode input, calculated as (Va + Vb)/2, yielding a common-mode gain
of unity.
Consider now a differential-mode voltage input to the circuit. A resultant current, i, will flow
through the series path defined by R1, R2 and R3. Specifically, Voutb - Vouta = i(Ri + R2
+ R3) by Ohm's law. The input voltage is similarly calculated as (V1 - V2)/R1 = i, or from
Figure 17:
Active ground circuit used for common-mode noise reduction. The common-mode signal at the input electrodes is inverted and
fed back through a current-limiting resistor (for subject projection). This circuit is particularly useful in reducing prevalent 60 Hz noise, which is
capacitively coupled into both signal input wires as described above.
the virtual-short property of opamps, (Vb - Va)/R1 = i. Therefore the differential-mode gain
of the first two opamps is (R1+R2+R3)/R1, which is always greater than unity. Therefore
the first stage of the instrumentation amplifier has the effect of amplifying differential-mode
signals while keeping common-mode signals at unity gain. The differential signal output of
the first stage just described is then fed into the second stage of the instrumentation amplifier,
a single opamp, which has a fixed CMRR determined by its transistor implementation. As a
system, the instrumentation amplifier therefore has a higher CMRR than the single opamp.
While an instrumentation amplifier offers a theoretical (albeit a second-order) common-mode
noise reduction over a single-opamp amplification stage, the hardware discussed in this paper
does not utilize an instrumentation aplifier. Pre-amplification, as described in section 3.1.5
below, utilizes a monolithic optical isolation amplifier that already has a reasonably high
CMRR (90 dB @ 60 Hz).
3.1.4
Increasing Input Impedance: CMOS Buffer Stage
High input impedence is requisite in a biomedical instrumentation design, as the signals of
interest (particularly electrophysiological signals) are extremely weak (on the order of several
hundred microvolts) and consequently cannot supply substantial current. Inherent in a
CMOS circuit is an extremely high input impedance and corresponding power amplification,
so a CMOS preamplifier opamp circuit serves as an ideal decoupling stage between the weak
Figure 18:
An instrumentation amplifier, which consists of three opamps, has an increased CMRR.
electrophysiological signal and subsequent analog signal processing circuitry.
The first stage of signal conditioning in the data acquisition hardware is a pair of CMOS
opamp buffers, one for the positive electrode input and one for the negative electrode. A
buffer is an opamp configured to have unity gain but an extremely high input impedance. In
figure 19, a schematic of the initial signal conditioning stage of the data acquisition hardware,
two buffers are shown preceding another amplifier. Note that, while the figure resembles the
instrumentation amplifier shown in figure 18, it is not in fact an instrumentation amplifier
since no resistor connects the negative terminals of the two input opamps. Consequently the
first stage of this circuit simply provides buffering and no improvement in CMRR.
However, high input impedance is itself a contributing factor to improved CMRR. Consider
again the circuit in figure 16, where R1 and R2 represent the input impedances seen at the
positive and negative terminals of the opamp, respectively, while R3 and R4 represent the
skin/electrode impedances at the positive and negative terminals, respectively. If we consider
placing a common-mode voltage source at the input (that is, we short the open ends of R3
and R4 together and apply a voltage Vi,), then the differential voltage that will appear at
the input terminals of the opamp is found by a simple voltage divider relationship. Namely,
Vdiff = Vin [(R1/(R1+R3))-R2/(R2+R4)]. If RI and R2 are nearly infinite, then R3 and
R4 become insignificant in the preceding equation so that Vdiff is approximately equal to 0.
3._
Vb
R2
Voutb
+.
It
107
R6
2
Vouta R6
+%0
107
U2
-Wo
R7
107
R3
Figure 19:
The two (CMOS) opamps to the left provide buffering, i.e. extremely high input impedance, that decouples the weak physiological
voltage at the electrodes from subsequent circuitry stages. The latter opamp is an optically-isolated pre-amplifier, described in section 3.1.5 below.
3.1.5
Pre-Amplification
The latter amplifier in figure 19, which represents the pre-amplifier stage of the signal acquisition hardware, is recognized in a single monolithic circuit, the Texas Instruments 3650KG.
This circuit provides optical isolation, which ensures that an electrical surge within the instrumentation circuitry cannot electrocute the subject, and conversely a surge at the input
terminals will not damage instrumentation circuitry beyond the preamplifier.
More dangerous is the threat of 60-Hz leakage current through the subject, minimal amounts
of which (less than 100uA (11)) can induce lethal ventricular fibrillation in catheterized human subjects. (The strongest source of such currents originates from capacitive coupling
through the power supply to the grounded instrumentation chassis.) Because the mice of
interest in the Krieger Lab study have surgically-implanted electrodes, they, just as catheterized human subjects, are vulnerable to fibrillation at low levels of leakage current. Figure
20 shows that as long as the chassis that houses the ECG hardware is properly grounded,
the minimal resistance of the metal case to ground will lead most of the current to sink to
ground through this pathway. In figure 21, however, when the chassis is no longer grounded,
leakage current only has a pathway to ground through the mouse, which could cause fibrillation if current levels are sufficiently high. The optical isolation amplifier discussed in this
section provides a very high dielectric interruption, or equivalently a very small capacitance,
in series between the lead wire and instrumentation, protecting the subject from acting as
a pathway for leakage current to ground. As illustrated in figure 22, the optical isolator
prevents a chassis/ground disconnect from routing leakage current through the subject.
120V
@ 60 Hz
ECQ
Hardware
Figure 20:
When the instrumentation chassis is properly grounded, capacitively-coupled leakage current originating from the transformer
follows the pathway as shown by the dotted line. In this example the subject is safe solely because of proper grounding of the chassis.
120V
@ 60 Hz
ECG
Hardware
Figure 21:
In this example the chassis is no longer grounded, and the only pathway for leakage current is though the subject. Sufficiently
high levels of leakage current through the subject can cause lethal ventricular fibrillation.
120V
@ 60 Hz
*
Figure 22:
In this example the chassis ground is disconnected but there is no pathway for leakage current through the subject due to the
extremely high-dielectric interruption, such as that provided by optical isolation. In this case the subject is safe in spite of improper grounding of
the chassis.
The gain of the TI 3650KG preamplification stage is determined by the equation 106 /(R4 +
R5), where R4 and R5 are as shown in figure 19. The preamplifier gain should be sufficiently large to boost the signal-to-noise ratio (SNR) yet sufficiently small to prevent opamp
saturation, or "railing" of the output. Opamp saturation, while not necessarily harmful to
instrumentation circuitry, represents an undesirable nonlinear region in the transfer function
of the circuit.
The physiological voltages produced by the mouse heart are on the order of 1OOuV, and the
dynamic range of the preamplifier is +- 12V. Accounting for different half-cell potentials in
the electrodes that could produce a differential DC voltage as high as 1OOmV, I expected a
gain of 50 was appropriate for the preamplification stage. Upon implementing the circuit
with this gain, however, I detected that the circuit frequently saturated. The saturation
resulted from baseline drift (extremely low frequency changes in the signal offset) that was
larger than anticipated. With a gain of 25 the circuit provided adequate SNR and, upon
reaching steady-state, never appeared to saturate.
3.1.6
High-Pass Filtering
The output signal from the instrumentation amplifier is input to an 8th-order Bessel highpass
filter with a cutoff frequency of 0.1 Hz. This filter serves to remove the DC offset due to
half-cell potential differences in the electrodes as well as other low-frequency signal noise
(i.e. baseline drift). The choice of a Bessel transfer function is motivated by the fact that
it has optimal phase response. That is, it has the desirable property of near-constant group
delay, or equivalently negligible phase distortion. This optimality in phase response comes
at the price of decreased rolloff steepness in the transition band relative to other transfer
functions. Figures 23 and 24 show Bessel high-pass filter amplitude and phase characteristics
contrasted with those of other filter implementations of the same order and cutoff frequency.
(For the sake of simplicity the figures are based off of low-order filters, but the concept of
relative rolloff steepness vs. phase linearity is fairly imparted with such an example.)
3.1.7
Secondary Amplification
After passing through the high-pass filter, the signal is again amplified, this time by a gain
of 52. The second amplification stage further increases the SNR of the signal and boosts the
signal voltage to a range appropriate for sampling with an A/D converter with a dynamic
range of +-10v. The amplification circuit, as shown in figure 25, is a simple feedback opamp
network utilizing the familiar gain equation 1+Rf/Rin, where in this system Rf = 510k and
Rin = 10k, to provide the gain factor of 52.
The signal entering this amplification stage, in contrast with that entering the pre-amplification
stage, is not offset due to half-cell potential differences and baseline drift because of the pre-
Vout
(dB)
.01
0.1
Frequency (Hz)
1
Figure 23:
The Bessel filter shows a slower rolloff than its counterparts, which is a slightly undesirable characteristic that is compensated
for by oversampling in the A/D conversion stage (see section 3.1.8).
90
4
hebyshev
Bessel
utterwor
0.
Frequency (Hz)
Phase
Figure 24:
The Bessel filter has a highly linear phase response, or equivantly a near-constant group delay, which means that the filter delays
all frequencies by the same amount. This linearity in phase means the filter has little to no phase distortion, a requisite feature for high-fidelity
signal acquisition.
ceding high-pass filter stage. As such this amplification stage can comfortably amplify the
signal by the rather sizable factor of 52 without saturating the amplifiers.
3.1.8
Low-Pass Filtering and Oversampling
Because the mouse ECG spectrum occupies 0-1kHz (9), the Nyquist sampling criterion
mandates that, with an ideal low-pass filter with passband 0-1kHz one should sample the
signal at 2kHz to avoid aliasing. Since a circuit filter implementation is never ideal, one
must enforce a relationship between the filter type, the filter's cutoff frequency, and the
A/D sampling rate that produces an acceptably small amount of aliasing. The filter transfer
function for the low-pass filter, as for the high-pass filter, was chosen to have a Bessel transfer
USA
Vin
LT1365
Vout
RR7
510k
R8
10k
Figure 25:
A noninverting negative-feedback opamp with gain 52 (determined by the ratio of the 510k to 10k resistors) serves as the final
amplification stage in the signal conditioning pipeline of the data acquisition hardware.
characteristic to minimize phase distortion. This optimization for phase response comes at
the expense of a slow rolloff in the transition region.
Oversampling is a technique often employed in systems using an antialiasing filter with
relatively slow rolloff. Though a derivation is not provided here, it can be shown that
sampling of an analog signal produces spectral copies of the analog spectrum at multiples
of the sampling frequency, denoted Fs, in the discrete-time frequency domain (8). Consider
the example in which a signal is filtered in the analog domain with a non-ideal lowpass filter
of cutoff 1kHz then sampled at 2kHz. The result is an aliased signal, which is manifested
in overlapping spectral regions in figure 26. One might consider building a higher-order
analog filter to reduce the transition band, which would prove costly and time-consuming,
to mitigate the effects of aliasing. Alternatively, one could sample the signal at a faster rate,
thereby spreading the spectral copies further apart, as figure 27 illustrates.
-4k
-3k
-2k
-1k
0
1k
2k
3k
4k
F (Hz)
Figure 26:
A signal bandlimited to 1.75kHz (due to slow rolloff of 1kHz cutoff Bessel antialiasing analog filter) sampled at 2kHz has spectral
copies at multiples of 2kHz and suffers aliasing (overlapping regions).
-10k
0.
...
10k
F (Hz)
Figure 27:
The same signal, bandlimited to 1.75kHz (due to slow rolloff of 1kHz cutoff Bessel antialiasing analog filter) sampled at 10kHz
(5x oversampling) has spectral copies repeating at multiples of 10kHz and does not suffer aliasing.
By then employing a high-order digital filter, which is cheap and relatively simple to implement, to lowpass filter the digitized signal, then decimating the signal to reduce the number
of time-domain samples required to store the signal, one could achieve the same effects as
with a high-order antialising filter without the hardware complexity!
F,_
-4k
-3k
I V
-2k
-1k
0
lk
I2k
I
3k
4k
F (Hz)
Figure 28:
After applying a high-order digital low-pass filter then decimating by a factor of 5, the spectra are spaced by multiples of 2kHz
but with no aliasing. This combination of digital lowpass filtering and decimation prevents the need for an expensive analog antialiasing filter.
Such a technique, known as oversampling, is often employed in data-sampling systems to
minimize the complexity and cost of analog circuitry and harness the power of fast digital
processing power. In the case of this system a high-order antialiasing filter is in place, but
since it is optimized for minimal phase distortion its rolloff is similar to that of a lowerorder filter. Thus this system employs 5-times oversampling, meaning that the signal is
sampled at 5 times the Nyquist rate of 2kHz. After being sampled at this high rate the
signal is then digally lowpass-filtered then decimated by a factor of 5 to give an effective
sampling rate of 2kHz. The initial use of oversampling minimizes aliasing, and subsequent
downsampling (after digital lowpass filtering) provides the minimum allowable lossless data
storage requirement without resorting to compression. Figure 28 shows the spectral content
of a signal after it has been oversampled (5x), digitally lowpass-filtered, then decimated by
a factor of 5. The spectral copies do not overlap, indicating that no aliasing has occurred.
However, the spectra are closely spaced, indicating that the signal is not oversampled and
therefore no disk storage is wasted when the data is archived.
3.2
Server Software
The Krieger lab currently uses proprietary software to acquire four channels of single-lead
mouse ECG data simultaneously. Several aspects of the software limit its power for our
intended goals. Specifically, the sampling rate is limited to 500 Hz, which is insufficient
for mouse ECG. The QRS complex, which contains high-frequency signal content, cannot
be captured adequately by such a low sampling rate. The undersampling manifests itself
in the time domain as variable QRS amplitudes since slow "snapshots" in time will not
consistently catch the fast-moving QRS complex at its apex. Additionally, the data collection
process is not automated. A user must manually save data every 12 hours, which stops all
four channels of data acquisition, in order to prevent the software from overflowing its
allocated buffer space. Furthermore, because the software is proprietary it is not amenable
to customization. We would like to network the software for remote monitoring of ECG
data, but its closed-source nature precludes tampering. Finally, the software is limited to
four channels of acquisition because it is bundled with proprietary hardware. As such, the
lab requires two computers to accomodate 8 single-lead ECG signals.
The server data acquisition software for this system runs on a RedHat Linux 8.0 RAID server
and was written in ANSII C, using a combination of custom-written code and C source
code libraries rovided by Microstar® Laboratories. Utilizing a data acquisition card made
by Microstar
Laboratories, it acquires 1 lead of ECG data on 16 mice simultaneously,
sampling each channel at a rate of 2kHz 2 . The software simultaneously streams the data
to disk and to a network device, permitting real-time network monitoring of the ECG (see
Section 3.3: Client Monitoring Software). The streaming nature of the design, combined
with the fact that the software runs on a high-capacity RAID server, allows full automation
of data collection for the duration of a mouse's lifespan (less than 8 weeks), precluding the
need for twice-daily manual file saving. The user will simply attach the requisite leads to
the mouse electrodes and start acquisition via an intuitive graphical user interface (GUI)
software package (see Section 3.2.1: Server User Interface Software).
3.2.1
Server User Interface Software
Most details of the data acquisition software described above are irrelevant to one who simply
wishes to begin collecting ECG data from a mouse. Therefore an intuitive graphical user
interface, written in HTML, SSI (Server Side Includes 3 ), Javascript, and CGI/Perl, allows
the operator (most likely a biology lab technician) to view the system's status as well as
to start and stop data acquisition on each of the 16 channels independently. A screenshot
of one portion of the interface software, that which provides a description of each channel's
status, is shown in figure 29.
Channel Nu ber
Status
Mouse NaIie
12
2
As detailed in section 3.1.8, the signal will initially be oversampled at 10kHz, then digitally lowpassfiltered and decimated by a factor of 5 to an effective sampling rate of 2kHz.
I
In an effort to enforce good record-keeping, the software requires the operator to enter
relevant information about the mouse prior to starting acquisition. The default mouse record
template requires that, at a bare minimum, the operator provide a mouse's eartag ID,
birthdate, and hardware channel to which the mouse is connected before starting acquisition,
as it is the unique combination of these fields that ensures a unique record name for each
mouse. After the user has entered this information, acquisition is fully automated until the
mouse dies and/or the user chooses to end data collection. Figure 30 shows the default
record template that a user must fill out prior to beginning data acquisition.
Figure 30:
The server operator, presumably a biology lab technician for this system, must fill out a record before data acquisition will begin.
This process enforces good record-keeping and, since the record information is stored in a back-end SQL database, will allow for simple web-based
search and retrieval of records.
Failure to enter reasonable data for one or more of these fields leads to an error message
that forces the user to comply, as is shown in figure 31. This form-checking functionality is
provided by a Javascript routine.
Figure 31:
The server operator entered an invalid month (13), and is therefore required to correct the error.
When the Javascript routine has verified that all form entries are reasonable, it then submits
the form to a back-end CGI script for processing. If the mouse record is unique then data
acquisition begins, while if the record already exists in the SQL database the user is warned
accordingly, as shown in figure 32.
3
The Apache web server, a product of the Apache Software Foundation, is the most widely-used web
Figure 32:
The user has attempted to enter a record into the database that already exists.
corrects the mistake.
Data acquisition is not started until the user
While the software should enforce proper formatting of required record fields, it should also
allow a certain degree of flexibility so that additional fields may be entered into the record
template to identify and differentiate mice in various studies. If the existing record template
fields are inadequate, presumably because the lab has introduced a new option (i.e. drug
regimen), then the user may add new template fields. Any new fields introduced are saved to
the record template and become requisite entries for each subsequent acquisition. Figure 33
shows the software as a new field is being added. A Javascript routine ensures that entries
only contain alphanumeric characters, lest the database contain corrupt field entries.
Figure 33:
In this screenshot the user is adding a field entry to the default mouse record template. Note that the field can be of type "text"
or of type "option." If the field is of type "text," then the user can freely type alphanumeric characters (enforced by a Javascript routine) when
creating the mouse record. If the field is of type "option," then the user will select one of several pre-defined mutually exclusive options from
which to select in the mouse record. Defining the options themselves for a given field is done in a manner similar to that of adding a field.
Just as a user should be able to add data fields, so should he or she be able to remove them.
server in the world. In addition to serving standard static HTTP content to clients, it can provide dynamic
content via its Server Side Include feature. An SSI statement embedded in an HTML document forces the
Apache server to run a designated script that determines dynamically the HTTP content to be provided to
the client. In the system outlined in this paper, SSI functionality is vital, as much of the web content is
dynamic. The screenshot in figure 26, for example, provides a readout of the current channel status by using
SSI.
Deletion of a field from the record template warrants a blatant warning that indicates to the
user that the action is permanent and a request for an administrator password 4 , as figure 34
shows:
Figure 34:
The user is attempting to delete a database field, which not only affects the record template but also removes this corresponding
field from the back-end SQL database. Such a drastic action should be verified! The user is required to enter an administrator password to proceed.
When the lab technician begins acquisition by filling out and submitting a record form
for the subject, the information is stored in a back-end SQL database for easy web-based
searching by specific criteria. This database contains a history of every subject entered into
the system, any template entries that were in existence at the time acquisition began, a
timestamp of when data acquisition began and ended, and the hardware channel on which
the data was acquired. Because an SQL database is highly malleable, it will be easy to add
analysis results to the database as they become available. Such flexibility is, in the author's
optimistic viewpoint, one means of making a "future-proof" system. For example, when
the ECG data is eventually analyzed for arrhythmia and ST elevation, the SQL back-end
architecture provides an easy means by which to provide any information related to the
analysis, such as pointers to annotation files, algorithms used for analysis, dates of analysis,
summary statistics of the analysis, etc.
3.2.2
Server Data Acquisition Software
Prior to beginning data acquisition per the process described in section 3.2.1 above, the
Microstar© driver and associated control software must be downloaded to the A/D card
and system daemons controlling data streaming to disk and the network must be initialized.
4
The administrator password referred to here is not the Linux operating system administrator password but rather a password associated with the administrator of the software. Therefore technicians and
researchers utilizing the software are not necessarily empowered to perform administrative tasks on the
computer itself.
Figure 35 shows the reminder message provided to the user when he/she attempts to begin
data acquisition. The Microstar® control software, written in C, and system daemons
controlling data flow to disk and the network, written in Perl, are all started with a simple
click of an icon, as shown in figure 36.
SBe sure Hermes Data Acquisition software is running.
Figure 35:
When a technician tries to begin acquiring data by using the GUI interface software (desribed above in section 3.2.1), he or she
is reminded to ensure that the necessary data acquisition software has been loaded.
3.2.3
WFDB Formatting, Server Self-Diagnostics and Email Notification of Errors
Because the user interface software employs a fairly complex interaction among many lengthy
HTML, SSI (Server-Side Includes), JavaScript, and CGI/Perl files, every action is carefully
verified to ensure that it does not produce unexpected results. Any error generated by a script
will produce an email with specifics of the problem so that the webmaster can immediately
re-activate the server and troubleshoot the bug that led to the error.
The server disk-streaming software operates on the premise that the ECG data being gathered is extremely important and the Krieger Lab must, in the unfortunate event of data
corruption disk errors, learn of the problem immediately. Thus 10 minutes of data are gathered on each channel and are saved to a file before a new file for the data stream is opened for
the next 10-minute segment. Software daemons continually monitor the streaming process
to ensure that the operating system (Linux) does not report an input-output (I/O) error at
any point. In the unlikely event of an I/O error, the software's error-checking mechanism
detects the error and sends an email to the system's webmaster with specific details about
the type and time of the error as well as which script generated the error. In the event that
there is such an error the system will stop collecting data, but since the files are broken into
10-minute data segments all previously saved files are unaffected by the error.
Dr. Roger Mark's group has, over the years, developed extensive software tools for the
analysis of ECG data. These analysis tools require that that the ECG data be stored in a
specific format, namely WFDB format (see http://www.physionet.org), and that a WFDBspecific header file be associated with each file. So that this software system produces data
that is amenable to analysis with WFDB tools, all files are converted in real-time into WFDB
format, and an associated header file is created by the disk-streaming daemon at the end of
each 10-minute cycle.
Figure 36:
Data acquisition software consists of C-compiled binaries that control the Microstar® A/D card as well as Perl system daemons
that control data flow to disk and to the the network. One click of an icon begins all these necessary processes.
3.2.4
Server Multicast Network Software
A Perl network daemon runs on the server to stream the WFDB-formatted data to a network
interface, allowing remote computers (on the MIT network) to monitor the ECG in real time
(see Section 3.3: Client Monitoring Software). The server has a static multicast IP address,
the first assigned to a non-system-administrator at MIT, that allows the server to broadcast
data to multiple clients simultaneously.
Multicasting is not a new concept, but it is relatively new and certainly less utilized in
the realm of computer networking when compared with the unicast architecture. A simple
analogy clearly distinguishes multicast from unicast networking: multicast is to radio as
unicast is to telephone. A unicast network, as its name implies, connects a single computer
to another computer. Such is the nature of TCP, arguably the best-known internet protocol,
which is a connection-oriented transport protocol for establishing network communication
between only two machines. Multicast connections, on the contrary, allow a server to stream
data from a given IP address, and any clients that wish to listen to the broadcast IP packets
may do so. Typically multicast software relies on a protocol known as UDP rather than on
TCP, as UDP does not perform handshaking nor allow the client any flow control (7). That
is, the server simply streams the UDP data, and if the client cannot process the data as fast
as the server sends it, or if the client detects that packets were lost in transit, the client has
no recourse.
The choice of a multicast architecture is more than an exercise in obscure network programming; rather, the choice of multicast over unicast enables a highly scalable architecture.
A unicast server would generate network traffic in direct proportionality to the number of
clients requesting data. A multicast server generates a constant amount of network traffic,
irrespective of the quantity of clients.
This system is intended to provide real-time analysis results to the client software, not just
raw ECG data. As future work leads to algorithms for arrhythmia detection and ST-segment
analysis, these reults can easily be streamed on separate network sockets for the client to
display along with the raw data. Because a single multicast IP address can accommodate a
multitude of independent connections, or sockets, such additions to server functionality will
not require that the server register an additional multicast IP address. In fact in its current
incarnation the server employs 16 separate sockets, all within the same multicast IP address,
one for each of the 16 hardware channels.
3.2.5
Server Security
Because the software that controls data acquisition is HTML-based and the server is online,
the software must enforce a simple security check that prevents malicious or unwitting users
from tampering with the lab acquisition processes from a remote computer. A CGI script
verifies that the IP address generating data acquisition begin and end operations is that of
the server itself, meaning that only users physically accessing the data acquisition server in
the mouse facilities may control data acquisition.
Another layer of security is included for archived data. The data should be accessible to offsite collaborators under controlled circumstances. Using Apache Web Server's secure HTML
access control feature, the software requires a username and password for access to mouse
records and data.
3.3
Client Monitoring Software
The server will continuously stream (multicast) real-time one-lead ECG data for each mouse
to its network interface. Client software monitors this streaming data and provides real-time
plots of the ECG. For the same reason that the server multicast software employs Perl,
so does the client software: Perl provides extensive and well-tested socket programming
Figure 37:
An example of the first-version monitoring software in action. This screenshot is taken while simultaneously viewing two archived
mouse ECG files. The first version was intended to allow viewing of 4 channels simultaneously, but visualization of 4 ECG signals at about 600
beats per minute would allow only a fleeting glance of data per channel within a given time period.
modules. The graphical user interface (GUI) aspect of the software employs Perl/Tk, a
specialized Perl module for incorporating graphics into a Perl script. The client software
displays a layered ECG strip that may show ECG data from either a static WFDB-formatted
file or from one of the available 16 network data streams.
By default the client starts in a scan mode, in which it monitors each channel (1 through
16) for 2 seconds before recycling. The user may opt to focus on a single channel or view an
archived ECG file as shown in figure 37. In a newer version of the client monitoring software,
shown in figure 38, the user may also choose to manipulate the amplitude scale, the number
of stripcharts stacked vertically on the screen, or the time scale. This second version utilizes
stacked stripcharts for a single channel rather than attempting to display four separate
stripcharts simultaneously, allowing the user to visualize a much longer data segment within
a single window. The second version's default behavior is to scan continuously each of the
16 channels for 2 seconds each, though a user may easily select to view a single channel or
view archived files.
4
Conclusions and Future Work
The reliable tools I have developed form the foundation for my subsequent research in the
area of intelligent cardiac patient monitoring. This robust data collection system comprised
of high-fidelity hardware and data acquisition/archiving software, in conjunction with networked software allowing real-time viewing of ECG plots, will facilitate much more effecient
research for the Krieger Lab and its associates. However, my ultimate goal, even beyond
Figure
38:
Version 2 utilizes
An example of the second-version monitoring software in action. This screenshot is taken from an EGG of a human suhject.
stacked stripcharts to allow viewing of a longer period of data within a given time period. Instead of providing for visualization
of 4 channels simultaneously, this version has a scan mode in which it cycles through each of the 16 channels for an adjustable period of time,
allowing the user to get a view of activity on all channels in a relatively short amount of time.
the scope of this thesis work, lies in a desire to use the system as a platform for writing
algorithms for the automated detection of arrhythmias and extraction of ST segment information. With such real-time capabilities in place, the K rieger lab could very effectively
realize its need to classify statistically the severity of heart disease in various genetic models.
Furthermore, they could begin to realize the more lofty goal of relating electrophysiology to
histology. Such an effort would be possible by utilizing the real-time system as an alert to
cardiac phenomena of interest, at which time they could sacrifice the subject for histology
studies.
Though the new system outlined in this paper represents a substantial improvement over
the existing data acquisition/archiving capabilities currently available to the Krieger Lab,
I see the work as a mere introduction to abundant possibilities for future developments. A
substantial base of knowledge and tools already exist in Professor Mark's group, a base to
which I hope to add. Specifically, George Moody's WFDB Project (part of Physionet) has
provided a widely-used standard for formatting and annotating data, as well as a substantial
number of algorithms for processing and viewing data and analysis results. With this standard in place my future goals of performing real-time analysis have a predefined impetus.
By providing real-time annotation outputs in WFDB format I can simultaneously create an
abundance of records that can contribute to the Physionet database (www.physionet.org)
while adding a real-time algorithmic component that serves as a logical evolution from existing retrospective analysis tools.
5
Acknowledgements
I would first and foremost like to thank Dr. Roger Mark for accepting me into his group
and for introducing me to the Krieger Lab Project. I have thoroughly enjoyed the intellect,
wit, and good humor of the great people who comprise his lab. I am particularly grateful to
Thomas Heldt and Dr. Wei Zong, officemates and friends, for consulting on subjects ranging
from poker to aesthetics of ECG instrumentation hardware. Wei Zong was particularly
helpful in explaining the details of active grounding circuitry as well as microshock scenarios.
I am also indebted to Dr. Monty Krieger for providing a challenging problem that will
continue to keep me busy for several years to come. His willingness to provide extensive
funding and complete trust in me has been indelible.
I would like to thank Steve Burns for his advice on 60-Hz noise reduction using active
grounding circuitry and his recommendation that I use DC-brushless fans to minimize 60-Hz
noise in the ECG hardware chassis.
Any acknowledgement would not be complete without thanking Dad, whom I miss terribly
and remember fondly, for providing a standard to which I will always aspire. As David
Flores, a San Antonio Express-News columnist, wrote of him, "If it's true that nothing is as
strong as gentleness and nothing so gentle as real strength, then Oefinger was the epitome of
both. That's why Nature probably stood up and gave him a standing ovation when he died,
applauding his good work on Earth and service to others during a life that ended much too
soon."
Most importantly I would like to thank my wife, Amy, who first encouraged me to follow
my lifelong dream of coming to MIT. Words at the end of a thesis can't possibly convey
adequate gratitude for this dream that she helps me to fulfill every day.
6
References
1.) Krieger, M. (Apr. 1998) Proc. Natl. Acad. Sci. USA, vol. 95 4077
2.) 2002 Heart and Stroke Statistical Update (2002) (American Heart Association, Dallas,
TX)
3.) Acton, S., Rigotti, A., Landschulz, K., Xu, S., Hobbs, H.H. & Krieger, M. (Jan. 26,
1996) Science 271, 518-520
4.) Braun, A., Trigatti, B.L., Post, M.J., Sato, K., Simons, M., Edelberg, J.M., Rosenberg,
R.D., Schrenzel, M., Krieger, M. (Feb. 2002) American Heart Assoc. Circulation Research
270, 271-275
5.) Dubin, Dale (Cover Publishing Co., 2000) Rapid Interpretation of EKG's 97-202
6.) Barkakati, Naba (IDG Books, 1996) Linux Secrets 183
7.) Hunt, Craig (O'Reilly, 1992) TCP/IP Network Administration 8
8.) Oppenheim, A.V, Schafer, R.W. (Prentice-Hall, 1999) Discrete-Time Signal Processing
196
9.) Ai, HB, Cui, XY, Tang., L., Zhu, WP, Ning, XB, Yang, XX (Oct. 1996) Sheng Li Xue
Bao: Studies on the Time Domain and Power Spectrum of High Frequency ECG in Normal
Mice 512
10.) Huhta, James C. and Webster, John G. (Mar. 1973) IEEE Trans. Biomed. Eng., vol.
BME-20 NO. 2 91-100
11.) Feinberg, Barry N. (Prentice-Hall, 1986) Applied Clinical Engineering459-465