ECG signal, the second-class presents the SNR classified ECG signals. We used 142
records in the training process, where 69 records are classified as SNR signals, 28
records are classified as AFIB signals and 45 records are classified as noisy ECG
signals, yielding a value of 100% of sensitivity and specificity. Moreover, for the
validation stage, we used 30 records where 24 records are well-classified achieving
accuracy value of 80%. Accordingly, test process achieved also a value of 73.3, where
22 records with 3600 samples for each record are well assigned to the right class, and 8
records are misclassified (3 records for the first class, 2 records for the second class and
3 records for the third class). These results are achieved due to the proposed architecture with 10 hidden layers, 10 neurons for each layer, with a sigmoid activation
function and finally Softmax classifier proves its ability not seen before by the system.
However, using a small number of hidden layers, the accuracy classification is with
82.6% and 87% using a high number of layers. Thus, the choice of a precision number
of hidden layers for getting precise number of parameters shows its worth in the
ECG-classification process.
Fig. 8. Confusion Matrix: training, validation and testing results with ANN1
108
L. Khriji et al.