Page 58 - AIH-1-4

Artificial Intelligence in Health A fuzzy system for heartbeat classification

selecting which records to use in the algorithm is based are plotted to verify whether the features of Table 1 are
on the number of specific beats in that record. Table 1 extracted correctly from the annotation function or not.
shows the input properties by input numbers that are Here, histograms of the other five types of heartbeats are
important for discussing the FIS membership functions. avoided, and only diagrams subject to the normal heart
For diminutive clusters, for each input, a cluster (Gaussian rate of one of the records are given in Figure 8. Histograms
membership functions) is created and the diminutive are helpful to see the data distribution and to show the
clustering algorithm normalizes the input properties. differences in the outputs. Histograms of Figure 8E-K are
The normalization layer of the ANFIS normalizes the approximately normal, while Figure 8A-D, and L are nearly
network weights. In Table 1, the left column shows right-skewed distributions. As the features are different in
different heartbeats, NSR, LBBB, RBBB, PVC, APC, and terms of measurement units, the horizontal axis differs in
PB. The first row lists extracted features. NF means no these histograms. For the histogram of the extracted QRS
information is available about that case. Six different heart feature, the center of data is located in 0.1 s, and most bins
conditions have their specific values in terms of temporal are devoted to this histogram. ST intervals, QT intervals,
and amplitude characteristics. Amplitude features are in and ST segments have a peak of about 50 bins. The peaks
millivolts, and temporal features are in both seconds and of P amplitudes, S offset amplitudes, and RR ratios are
milliseconds. about 40 bins. PR intervals, P wave intervals, and Q onset
For NSR, both RRs and RRp intervals are the same, so amplitudes have approximately the same height (about 100
their ratio is equal to 1; here, the PR interval is longer than bins). PR segments and R amplitudes reach 60 bins. The
the QRS interval. In LBBB, RBBB, PVC, and paced, the QRS centers of histograms in Figure 8B, C, I, J, and K are between
interval is the same and it is more than 120 ms. There is no 0 and 0.5 s. Q onset and S offset amplitudes have negative
specific information about the PR interval of LBBB, RBBB, amplitude in the duration of [−1 0] millivolt. However,
PVC, and APC. APC and PB are in common in terms of S offset amplitudes are higher than Q onset amplitudes.
R amplitude. It seems that the ST segment does not play Among these extracted features, seven of them are more
much significant role in classification because, in four efficient to go through ANFIS as inputs for classification
heart conditions, there is no sign of this feature; however, it effectively. The selected features for this research are PR
improves the simulated results. RR features (RRs, RRp, and segment, P wave interval, P amplitude, Q onset amplitude,
RRs/RRp) are not prominent factors in PB. ECG properties S offset amplitude, T wave interval, QT interval, and ST
are extracted in 30 min for about 646400 samples. Figure 7 interval. The subtractive clustering generates an initial
shows the results of feature extraction algorithms on the FIS for each ANFIS. Then, the central parameters of the
recordings of the ECG on which the heart condition is to Gaussian function and standard deviation are adapted by
be diagnosed. These signals can be seen from a modified the ANFIS. The number of fuzzy membership functions
limb lead II with 360 samples per second of approximately affects the ANFIS model; fewer membership functions
646,400 samples. Six types of heartbeats are labeled as cause less complexity and a lesser run-time. In Figure 9,
“N,” “L,” “R,” “V,” “A,” and “/” to represent the various the initial FIS for NSR ANFIS is shown.
conditions. At this stage, all types of beats are identified, When Gaussian membership functions are used, the
along with features such as the P-wave, QRS complex, and transitions between membership values are smooth and
T-wave, including their onset and offset. The duration of continuous. The model can detect minute variations since
this record is long and the detail of these annotations is not the input data are smooth, which is particularly useful for
clear enough. Hence, it is better to separate one random detecting variations in heartbeat signals that can result
heartbeat as below. Histograms of NSR characteristics from noise or heart rate variability, among other factors.

Table 1. The input feature range of an electrocardiogram signal 21
ANFIS QRS interval (ms) PR interval (ms) R amplitude (mV) ST segment (ms) RRp interval (s) RRs interval (s) RRs/RRp
NSR 80 – 100 120 – 200 1.5 – 2 80 – 120 0.6 – 1.2 0.6 – 1.2 1
LBBB >120 NF NF NF NF NF NF
RBBB >120 NF NF >120 NF NF NF
PVC >120 NF <2 NF <0.6 >1.2 >1
APC <80 NF >2 NF <0.6 >1.2 >1
PB >120 >280 >2 NF NF NF NF
Abbreviations: ANFIS: Adaptive neuro-fuzzy inference system; APC: Atrial premature condition; FN: False negative; FP: False positive; LBBB: Left
bundle branch block; NF: No feature; NSR: Normal sinus rhythm; PB: Paced beat; PVC: Premature ventricular contraction.

Volume 1 Issue 4 (2024) 52 doi: 10.36922/aih.3367