Page 52 - AIH-1-4

Artificial Intelligence in Health A fuzzy system for heartbeat classification

Neural networks provide efficient ways for ECG eliminates the need for manual feature engineering by
classification without requiring pre-processing, such automatically extracting relevant features from the input
14
as multilayer perceptron (MLP). Based on the conjoint data. This capability can save time and effort during the
15
use of the MLP that was trained by an enhanced particle preprocessing stage, enabling the model to effectively
swarm optimization algorithm, the ECG arrhythmias capture minute patterns in heartbeat signals that might not
were classified. Four types of heart rates were classified be immediately apparent to human observers (Equation I).
16
by identifying QRS features that were extracted from
multi-resolution wavelet transform. In addition, a − ( 1 z ) 2
17
−6
quality-aware mechanism was employed to classify ECG Hz () = 2 (I)
−1
beats, reducing false alarms and ensuring accuracy. A − ( 1 z )
18
multi-module neural network system was developed to
classify ECGs, specifically addressing the issue of heartbeat The amplitude response is given by Equation II, where
imbalance. A research architecture employs ANFIS to T is the period of sampling.
19
learn fuzzy logic, using inputs that are preprocessed with
2
the subtractive clustering method. Five morphological HWT) = sin(3 wT) (II)
(
and five statistical ECG features were used to classify the sin( wT /)2
2
patient’s heartbeats based on whether they were irregular
or normal. Six various heart conditions, including normal The low-pass filter from Equation I is represented by
20
sinus rhythm (NSR), PVC, atrial premature condition the difference equation shown in Equation III.
(APC), left bundle branch block (LBBB), right bundle y(nT) = 2y(nT-T)-y(nT-2T)+X(nT)-2X(nT-6T)+X(nT-12T)
branch block (RBBB), and paced beat (PB), were detected (III)
by an ANFIS. A weight assignment method based on
21
multi-label ECGs, combined with an ensemble classifier, For the low-pass filter, the cutoff frequency and gain are
was applied for classification. The neuro-fuzzy system has set at 11 Hz and 36, respectively, with a processing delay
22
proven helpful in disease diagnosis, while some research of six samples. A high-pass filter has also been designed,
23
has employed eigenvalues and DL for ECG classification. with its transfer function represented in Equation IV.
24
Considering previous research, some challenges remain, The amplitude response is given in Equation V, and the
such as accuracy improvement, complexity reduction, difference equation is provided in Equation VI. Here, the
speed increment, and power reduction. There is a growing low cutoff frequency is 5 Hz, the gain is 32, and the delay
interest in computer-aided identification and diagnosis is 16 samples.
of cardiac illness using ECG data. Some researchers are −16 z )
−32
−+
turning to neural networks to overcome the drawbacks Hz () = ( 132 z + (IV)
of manual feature selection methods. However, it is still + ( 1 z )
−1
difficult to build and choose a high-performing diagnostic
model that is appropriate for clinical implications. 25,26 26 + sin 2 (16wT )  1 2
(
3. Data and methods HWT) =   wT   (V)
cos  
To improve the speed and accuracy of ECG classification,  2 
a modified ANFIS structure is proposed (Figure 2). The
advantages of neural networks and fuzzy logic systems y(nT) = 32x(nT-16T)-[y(nT-T)+x(nT)-x(nT-32T)] (VI)
are combined in ANFIS. Its hybrid approach allows it After filtering, the five-point derivative, along with
to learn and comprehend complex patterns in the data the transfer function (Equation VII) and the amplitude
in an adaptive way, which greatly increases its versatility response (Equation VIII), is applied to differentiate the
for classification jobs. ANFIS can effectively handle this signal. Equation VII indicates the derivative operator.
uncertainty because of its fuzzy logic component, which The derivative procedure gives a large gain to the high-
allows for approximate reasoning and decision-making in frequency components resulting from the high slopes of
the face of ambiguity and vagueness. Over time, ANFIS the QRS complex while suppressing the low-frequency
models can adapt to changes in the input data distribution components of the P and T waves. The difference equation
or environment. This adaptability is especially useful for in Equation IX results in a nearly linear frequency response
recognizing heartbeats in real-world scenarios, where between DC and 30 Hz. Next, the signal is squared point
data features may vary due to factors, such as patient by point, as indicated in Equation X. This squaring
condition, activity level, or sensor positioning. ANFIS operation suppresses the small differences from the P and

Volume 1 Issue 4 (2024) 46 doi: 10.36922/aih.3367