REAL-TIME HEART FAILURE DETECTION BY

ANALYSIS OF EGM SIGNALS

REAL-TIME HEART FAILURE DETEC TION BY ANALYSIS OF EGM SIGNALS

M.Karthik Reddy K. Akhila

ECE III rd Year ECE III rd Year

[email protected] [email protected]

CVR COLLEGE OF ENGINEERING

ABSTRACT

Novel Methods of cardiac rhythm detection are proposed that are based on time-frequency
analysis by a weighted overlap add(WOLA) oversampled filterbank. C ardiac signals are obtained
from intracardiac electrograms and decomposed into the time-frequency domain and analyzed by
parallel peak detectors in selected frequency sub-bands. The coherence of the sub-band peaks is
analyzed and employed to detect an optimal peak sequence representing the beat locations.

Keywords: C ardiac Arrhythmia, Heart Failure alerts.

1. INTRODUCTION

The objective of this paper is rhythm classification and event detection based on the intracardiac
electro gram (EGM) signals. The proposed methods are designed for implantable devices that
should operate on extremely low power budgets. In the meantime, these methods should operate
in real time and the processing delay should be in the minimal range acceptable for such
applications. The detection methods should be very reliable and robust to interference, noise, and
morphology variations.

C urrent practical methods of cardiac rhythm detection employed in implantable cardioverter

defibrillators (IC Ds) are generally based on beat-by-beat time-domain analysis. Although
research has been conducted to exploit more sophisticated signal processing such as wavelet
transform and template matching for event detection, the new methods have rarely been
employed in practical systems due to their computational and power demands and issues related
to the reliability of their detection.

C urrent challenges in reliable rhythm detection for implantable cardiac rhythm management
(C RM) systems such as IC Ds are the following:

• Inappropriate device therapy (IDT) amount to a considerable rate (between 10 to 30%) in

various devices. IDTs occur due to low EGM signal quality, sinus tachycardia, supraventricular
tachycardia (SVT), myopotential interference, external interference, and Twave over sensing. IDT
is painful to the patient and depletes the device battery power more quickly. IDT is also
potentially harmful to the patient as it puts the patient at risk of device-induced VT (pro
arrhythmia) that might be dangerous and hard to detect by the IC D.

• Missing serious cardiac events compromises the reliability of the C RM devices. This happens
due to many reasons including quick morphology, rate, and even polarity changes of the EGM
signal, abnormally wide R-waves and P-waves, and external noises. The problems are aggravated
due to the fact that often patients have to simultaneously use medicines that alter the EGM
waveforms.

• Following a device therapy (low-energy pacing or high-energy cardio version), it is essential to

quickly redetect the EGM rhythm to analyze the effectiveness of the therapy. Quick redetection,
however, is tricky as often the device therapy polarizes the EGM electrodes or causes baseline
variations of the EGM signal (sometimes in form of low-frequency oscillations with magnitudes
larger than the EGM beat itself). Often a blackout period has to be applied before any reliable new
sensing.

• As increasingly sophisticated multichamber C RM devices become available offering the

physicians more choices in programming the devices, more powerful signal processing is required
to eff iciently handle the multichannel information while offering the physician less complicated
and reliable programming scenarios. Also, increased immunity to cross-channel interference is
necessary.

To cope with the mentioned problems and to improve the rhythm detection accuracy, we propose
a multitiered detection method that is based on time-frequency analysis of the EGM signals by a
weighted overlap-add (WOLA) filter bank. The WOLA filter bank can be efficiently implemented on
an ultra-low power platform targeted for real-time low-delay and implantable devices.

Methods based on time-frequency analysis have already been proposed for electrocardiogram
(EC G) and QRS detection. However, the proposed methods are not qualified for ultra-low power
implementation on implantable devices since they are either (a) too complex or unsuitable for
real-time applications or (b) specifically designed for (and evaluated on) EC G signals and do not
provide the robust and reliable performance essential for EGM signal processing.

In EC G detection, the EC G signal is analyzed by a critically sampled polyphase filterbank,

extracting six features that are all based on the accumulated energy (or absolute value) of groups
of subbands. Each energy feature is processed by a peak detector that compares the signal
moving average to a threshold that is determined based on an estimation of the background
noise. Peaks are then refined in a cascade of five stages (levels). In one stage, two different
thresholds are used to detect the peaks of the same feature. The peaks are then combined in
parallel. The method is evaluated on a standard EC G database with satisfactory performance.

The objective of this paper is to present a real-time detection method for cardiac event detection
using the intracardiac EGM signals. Sensed EGM signals differ from EC G signals in many aspects.
Major differences are:

1. EGM signals provide direct access to individual heart chambers, most importantly right
ventricular and right atrial, at the signal source. In contrast, EC G signals provide a
combined signal after propagation of various waves to the body surface.
2. Relative timing of various EGM signals is very important as it could be employed to
discriminate various cardiac events (e.g., SVT versus VT) accurately. For EC G signals,
however, such timing information is not available.

3. Unlike EC G signals, sensed

EGM signals are prone to cross talk. For example, far-field R-waves (FFRWs) might occur
when the much stronger ventricular signal interferes with sensing of the weaker atrial
signal. Although FFRWs are a major problem with unipolar electrodes, they interfere with
bipolar electrodes to a lesser degree.

Fig. 1. WOLA Over sampled Filter bank

Signal processing strategies are therefore greatly different for EGM signals as compared to EC G
signals. Rather than critically sampled polyphase filter banks, a very efficient WOLA over sampled
filter bank for time-frequency analysis of the EGM

signals is employed. Subband peaks are detected directly from the subband signals (absolute
values) by recursive averaging with no absolute thresholds. Subband peaks are combined in
parallel by exploiting the synchrony of the subband signals at the beat time. Furthermore, to cope
with the wide range of possible beat rates and morphologies of the EGM signal, the narrowband
(complex) subbands are merged to obtain wideband-subband signals to be used for wideband
event detection. The results of the wideband and narrowband detections are then combined for
robust detection.

As the intention here is to describe the basis for the detection method, attention is limited to
single-electrode analysis; extension to multiple-electrode analysis is straightforward. Also, the
algorithm simplicity is a major consideration in this case since we are targeting low-power, real-
time, and implantable applications.

This paper is organized as follows: Section 2 presents details of the detection algorithm, Section 3
discusses evaluation of the methods using the EGM signals in clean and in additive noise, and
Section 4 presents conclusions.

2. THE PROPOSED DETECTION METHOD

2.1. General

A time-domain EGM signal x ( n ) is analyzed by an oversampled filterbank (depicted in Figure 1)

that is efficiently implemented using a WOLA structure . The filterbank parameters, adjusted by
optimization for this application, are K = 32 subbands, analysis window length of L = 256, subband
decimation factor of R = 4, and oversampling factor of OS = K/R = 8.

Fig. 2. From top to bottom, a segment of ventricular EGM signal, 4 WOLA subband energies with
their average, maximum, instantaneous, and peak signals, their corresponding binary pulses and
the optimal detected pulses(bottom row)

At the output of WOLA analysis, K complex-valued subband signals are obtained: Z k ( m ), k = 0,

1, . . . , K - 1, where m is the subband time index. For real input signals, only half of the subbands
are stored and processed due to Hermitian symmetry. The subband time-index m is updated
every R = 4 input samples when a new block of WOLA subband signals is available. Subband
signals are then framed with a frame length of 3 seconds and a frame shift of 2 seconds. The
frame length should be chosen long enough to cover more than one beat for slow beats (around
60 beats per minute, bpm) and to provide enough beats for statistical analysis. At the same time,
the frame should be as short as possible to track the dynamics of the quickly varying beats. The
choice of frame shift is rather arbitrary and depends on how often a decision is needed. Notice
that irrespective of the frame length and frame shift, the WOLA analysis is continuously applied to
the input signal, yielding a new block of subband signals for every R input samples.

The cardiac beat is often represented by a sharp pulse in the EGM signal. As a result, the
magnitudes of subband signals ( | Z k ( m ) | , k = 0, 1, . . . , K - 1) exhibit mainly coherent
peaks at the time of cardiac depolarization. A major objective in this research is to exploit this
subband coherence (“synchrony”) between various subbands. Based on the synchrony analysis, a
final beat sequence (“optimal beat”) is detected for every frame as detailed in the next section.
Then the periodicity and the regularity of the optimal beat are combined with the synchrony
measure to detect the underlying cardiac event.

2.2. Subband peak detection and synchrony analysis

In the first stage, peaks are detected in selected subbands. Given the subband magnitude signal |
Z k ( m ) | for subband k , its maximum is tracked with a two-time-constant first order recursive
filter. C onsidering two filter coefficients of 0 . 9 < á m 1 < 1 and 0 . 1 < á m 2 < 0 . 5, the
following pseudocode describes how the maximum signal ( M k ( m )) is calculated:

(1)

After each peak, the filter acts as a leaky integrator. Accordingly, the first filter coefficient ( á m
1 ) is selected close to one. This controls the so-called “release time” of the filter. The other
coefficient is selected smaller for the filter to react quickly to the next peak. Similarly, the
average signal A k ( m ) is tracked by a two time- constant recursive filter with 0 . 7 < á a 1 < 1
and 0 . 5 < á a 2 < 0 . 7. At each time-instance m , a peak value ( P k ( m )) is detected by
comparing the three values of | Z k ( m ) | , M k ( m ), and A k ( m ) (the instantaneous,
maximum, and average values) as described in the following pseudocode:

P k ( m ) = 0,

If {| Z k ( m ) | > A k ( m-1 ) & | Z k ( m ) | > 0.5 M k ( m-1 )} &

{A k ( m-1 )< 0.9 M k ( m-1 )}, P k ( m ) = | Z k ( m ) |

(2)

Between two distant cardiac beats, it is possible that A k ( m ) and M k ( m ) converge to each
other. To prevent peak detection in this situation, the last term in the condition above is included.
Notice that no absolute threshold is used in the peak detection and only relative thresholds are
employed. By analyzing various beats in subband domain, it was observed that peaks appear
more distinctively in the first half of the subbands. As a result, peak detection is limited to
subbands 2–9 (out of 1–16). The first subband is ignored as it mostly captures noise and baseline
wander. Figure 2 illustrates a frame of atrial EGM signal (top graph), and four sets of WOLA
subband signals of | Z k ( m ) | , M k ( m ), A k ( m ), and P k ( m ) (instantaneous, maximum,
average, and peak).

Following the peak detection, each subband peak signal P k ( m ) is converted to a binary (0/1 for
peak/no-peak) signal B k ( m ). This greatly simplifies further processing. To embed more
robustness in the algorithm and to avoid detecting short-term spurious peaks, we search every
frame of binary peak signal B k ( m ) for a pattern of consecutive peaks (1s) followed by a block
of zeros (e.g., {1 1 1 0 0 0}). For every pattern found, the falling edge of the binary peak signal
is registered as a valid peak. The peak location is marked by a block of three 1s ({1 1 1}) and
the rest of the peak signal is reset to zero. Replacing the peak by a block of 1s (rather than a
single 1) increases robustness in the next stage of synchrony analysis. All further steps of
processing are applied to the binary peak signals B k ( m ).

2.3. Synchrony analysis and robust beat detection

In the next stage, the degree of synchrony between various subbands is measured by applying
simple AND operations to the binary peak signals. For each possible pair of signals B k ( m ) and
B l ( m ), k! = l , synchrony of the pair S k , l (in percentage) is calculated as follows:

S k , l = 100 NP(B k (m) & B l (m))

Max [NP (B k (m),NP(B l (m)] (3)

where function NP ( · ) denotes the number of peaks in a frame of binary peaks and & denotes
the logical AND operation. The synchrony is evaluated for all nonidentical pairs

(for 8 subbands this involves 28 AND operations on frame pairs). To minimize the effect of noise
and interference, only the top 3 synchrony scores are considered as measures of the frame
synchrony. The top 3 scores are compared to fixed synchrony thresholds to classify the frame of
subband beats as perfectly synchronous (Syn = 4), as borderline synchronous (Syn = 2), or
asynchronous (Syn = 0).

The top 3 binary pulse pairs are used to robustly detect the beat times. Applying a majority-voting
rule, beats are detected from the 3 pairs (after the logical AND operation within each pair) when 2
out of the 3 pairs exhibit simultaneous beats. C onsidering the peak extension to a block of three
1s, this method proved to be very robust when the signal quality was compromised by noise or
due to flutter and fibrillation. As a result of beat detection, an optimal beat sequence OB ( m ) is
obtained for every frame. Depicted in Figure 2 (rows 6–9) are binary subband pulses for the EGM
segment together with the optimal-detected beat (bottom row).

2.4. Analysis of periodicity and regularity

Once an optimal beat sequence OB ( m ) is obtained, it is analyzed to find the beat rate and the
regularity of the beats. A set of thresholds for periods of various cardiac events is used to set
histogram edges as [0, FibPer/2, FibPer, FlutPer+1, TachyPer +1, SRMax, infinity]; where FibPer,
FlutPer, TachyPer, and SRMax indicate the largest acceptable periodsfor fibrillation, flutter,
tachycardia, and sinus rhythm, respectively. The beats are classified in a period histogram with 6
bins specified by the above edges. The mode (bin index for the most populated bin) of the
histogram ( T m ) is an indicator of the periodicity. For the periods in the “acceptable” range of
(FibPer/2, SRMax), the mean period ( T ) and the standard deviation-to-mean ratio ( ó/ì ) are
calculated. If either of the T and T m fall in the fibrillation, flutter or tachycardia range, the period
zone indicator would be set to show the corresponding event. For sinus rhythm, however, both T
and Tm should indicate a sinus rhythm. In all cases, we chose to use the mean period T to find
the rate as beat per minute, bpm = 60 /T . So, it is possible that the period zone indicator show a
flutter since Tm is pointing to a flutter while average beat rate is still slightly below the minimum
flutter rate. ó/ì (of periods) is an indicator of the regularity; typically for very regular beats ó/ì <
20%, for very irregular beats ó/ì > 40%, and ó/ì values between the two ranges indicate
moderate regularity. In case unusual lack of EGM activity (longer than the slowest possible
rhythm) is detected within the frame, the irregularity flag is set (Ireg = 1).

2.5. Event detection based on subband features

The synchrony analysis provides both the optimal beat OB ( m ) and the synchrony score (Syn =
0, 2, 4). Based on these and the periodicity and regularity of the optimal beat ( T , T m , ó/ì , and
Ireg), cardiac events are classified as one of the following eight events:

(1) Stable sinus rhythm (SR),

(2) Transitional SR (T-SR),

(3) Stable tachycardia (VT or AT),

(4) Transitional tachycardia (T-VT or T-AT),

(5) Flutter (VFLUT or AFLUT),

(6) Fibrillation (VFIB or AFIB),

(7) Synchronous but irregular rhythm (Syn-Irg),

(8) Unclassified,

where VT, VFLUT, and VFIB represent ventricular events of tachycardia, flutter, and fibrillation,
respectively. Similarly, AT, AFLUT, and AFIB represent the corresponding atrial events. When the
mean period is within the range for sinus rhythm but the rhythm is irregular, a transitional event
of T-SR is detected. A similar criterion is used in detection of TVT or T-AT. Event (7) is detected
when the synchrony is perfect but periods are too irregular or insufficient in number to be
considered for other classes. Finally, event (8) is reserved for unclassified rhythms. A flowchart of
the event detection algorithm is depicted in Figure 3. As shown, the algorithm sets a series of
traps for various events. It first tries to identify fibrillation or flutter (classes {5, 6}). If none is
detected (state A in the figure),it searches for fast beats (classes {3, 4, 7}) and then sinus rhythm
(classes {1, 2, 7}). If none of the traps succeeds in detection, the beat remains unclassified (class
8).

2.6. Detection by wideband filterbank

The expected range of cardiac beat rates is very wide, from less than 50 bpm to over 300 bpm.
As a result of the classic time-frequency resolution trade-o . , the time resolution of the
narrowband filterbank (WOLA analysis with K = 32) is insufficient to separate two closely spaced
beats. The problem is compounded when the signal quality is further compromised during flutter
or fibrillation. An effective solution is to use a filterbank with wider subbands. In uniform
filterbanks, it is possible to merge the subbands through a simple postprocessing. Specifically, in
the WOLA filterbank, we can combine, for example, every neighboring pair of complex subband
signals to obtain a wideband analysis, doubling the time resolution. Since all of the subband
signals have baseband spectrums, to combine subbands one has to modulate the bands to line
them up sequentially. For merging two subbands, for example, one has to apply a complex
modulation to the higher-frequency subband and add the results with the lower-frequency one.

To achieve a higher temporal resolution,

the low frequency subbands in pairs (subbands 2–7) and in a group of four (subbands 2–5)
resulting in four new wider subband signals are combined. In merging subbands, the limiting
factor is the filterbank-oversampling factor (OS = K/R ). As the effective number of bands ( K )
decreases for wider subbands, the potential for aliasing increases. The aliasing is kept minimal
with the proposed WOLA setup ( K = 32,OS = 8), when grouping in pairs (equivalent to K = 16
bands) or in fours (effective K of 8) since the oversampling factor for the combined bands is at
least OS = 2. Aliasing and distortion are also greatly reduced by proper prototype filter design.
Using the four wider subbands, a wideband peak detection, synchrony, periodicity, and regularity
analysis similar to the narrowband case was performed.

Figure 3: Flowchart of the event detection algorithm

2.7. Low-frequency detection

The EGM wave morphology is very diverse. Among all various forms, there are cases where the
EGM beat lacks a clear strong impulse at the beat instance. Instead, a periodic waveform with
wide R-waves or P-waves (for the ventricular or atrial signals, resp.) is observed with weak
impulses at the beat locations. Detecting such beats is problematic in noise since the EGM
waveform, exhibiting a low-pass behavior, is presented mostly in very low frequency bands. To
increase noise robustness for such cases, a third method offbeat detection by using only the
peaks detected in subbands 2 and 3 is added. The synchrony between the two subbands as well
as the periodicity and regularity of optimal beat (AND result of the two) is calculated as before.
This is called low-frequency (LF) detection here.

2.8. Multitiered beat and event detection

Taking the narrowband filterbank detection as the default, the wideband system is selected when
all of the following conditions are met.

1. Wideband detection shows perfect synchrony (Syn = 4).

2. Wideband detection has ó/ì < 40% or less than the corresponding value for the
narrowband detection.

Switching to the LF detection occurs when all of the following conditions are met.

• Both the narrow and wideband systems are not synchronous (Syn < 4); or ó/ì of the LF
detection is superior to (less than) each of the other two systems by at least 40%.

• The LF system detects less than four pulses in the frame.

• The LF system is not detecting fibrillation or flutter.

3. SYSTEM PERFORMANCE EVALUATION

The EGM signals recorded with bipolar electrodes were utilized for evaluation since they provide
the best quality. The two bipolar signals were recorded from Right Ventricular Apex and High Right
Atrium, called RVAb and HRAb, respectively. All of the RVAb and HRAb signals were used for
system evaluation.

Fig. 4: Algorithm Performance for an HRAb EGM signal

The EGM signal was digitally decimated from the original sampling frequency of 1000 Hz down to
Fs = 250 Hz. Using a small subset of the library representing various events, the system
parameters were tuned until there were no missed events, and the frame-by-frame classification
was as accurate as possible. Then, for every case in the library, the detection algorithm was
tested and the results were compared to the physician-certified annotations. Figure 4 depicts a
typical output summary of the detection system showing various detected events and the beat
rate. Notice that there is no “training” involved in the proposed detection system; rather system
parameters had to be optimized on a subset of the EGM data.

3.1. Statistical performance evaluation

For practical applications, noise robustness is a very desired feature of any cardiac event
detection. For a statistical analysis, the performance in noise (or with reduced algorithms) against
the benchmark detection was compared.

To simulate various noise conditions, five different noises were added to the EGM signals:

1. White Gaussian noise, (0, ð ) band,

2. lowpass noise, (0, ð/ 4) band,
3. bandpass noise, ( ð/ 4, ð/ 2) band,
4. highpass noise, (3 ð/ 4, ð ) band,
5. Tonal 60 Hz noise.

Noises (2)–(4) were obtained by filtering white noise. To adjust the noise level for a given signal-
to-noise ratio (SNR), one needs to measure the EGM signal power. Due to the variability of the
EGM signal in terms of magnitude, polarity, and morphology (very wide waves to very sharp
ones), measurement of long-term power is inadequate. Instead, we adjusted the noise level
based on tracking the short-term (4 second) EGM signal envelope (rather than power).

To quantify the performance, in presence of noise the 8 cardiac events (Section 2.5) are grouped
into two separate groups, {1–4} (SR/Tachy) and {5-6} (Fib/Flut), and the detection performance
using frame counts of TP, TN, FP, and FN defined as follows was measured.

• TP, true positive: correct detection of {5-6}.

• TP, true negative: correctly not detecting {5, 6}.

• FP, false positive: falsely detecting {5, 6}.

• FN, false negative: falsely not detecting {5, 6}.

From these frame counts, the Fib/Flut positive predictivity (+ P ) and negative predictivity ( - P )
were calculated as follows:

Depicted in Table 1 are the total frame counts of TP, FP, TN, and FN for the five noise types. Also,
Table 2 summarizes the + P and - P measures for the five noises. As expected, the adverse
effects of white noise on subband detection are worse as they corrupt all the bands equally. On
the other extreme, the system shows immunity to even 0 dB SNR tonal (60 Hz) and highpass
noises. Our careful observations revealed that no block event was missed or misrecognized with
five noise types in Table 2. Moreover, most of the recognition errors in noise occurred before or
after fibrillation or flutter events when the quality of EGM signal was already compromised.

Table 1: Number of TP,FP,TN, and FN frames for five noise

types

Table 2: Positive and Negative

Predictivity for five noise types

4. FUTURE SCOPE

The current work is primarily useful in a hospital or monitored setup. The WOLA filterbank and the
complete setup can be converted to a IC form or strip form (that can be worn like a watch). The
input EGM can be transmitted to the strip using some non-invasive, high-speed technology like
RFID continuously. The required power for the transmitter can be supplied by the strip, which can
be recharged. The output i.e. event detected can further be made to be transmitted through a cell
phone to a predetermined emergency number. This can be useful in saving the lives of many
people.

5. CONCLUSIONS

The subband-based methods proposed in this paper for processing intracardiac EGM signals offer
a robust and reliable performance by employing parallel narrowband peak detectors. Proper and
efficient combination of subband peaks by synchrony analysis is a major milestone. The method,
demonstrates excellent performance in terms of accurate event detection and beat-rate
measurement even in fibrillation or flutter when the signal quality is compromised. Evaluation in
noise has also demonstrated significant robustness to noise. This method is simple enough for
implementation on an ultra-low power WOLA filterbank platform and requires only simple
operations as a result of using binary peak signals.

REFERENCES:

1. Biomedical Digital Signal Processing, Willis J. Tompkins, Prentice-Hall India

2. AMI Semiconductor

C ollected and C reated by youtrick.com .....