Automatic Arrhythmia Classification from Electrocardiogram Measurements with Deep Learning
Özet
ECG signals have an important place in detecting arrhythmias. Arrhythmias are irregular heartbeats. One of the most popular studies in this field is the classification of arrhythmias with artificial neural networks. In the thesis study, a classification study of arrhythmias was carried out with artificial neural networks using ECG lead signals. In this way, it can be determined directly whether there is an arrhythmia or not as soon as an ECG recording is taken.
The dataset used in the study was obtained by combining the ECG recordings in the PTB-XL and Chapman datasets. The types and numbers of arrhythmias in each data set vary. When working on a single data set, trained models will be successful in classifying certain arrhythmias. However, it will fail to classify arrhythmias that are not included in the dataset. To avoid this problem, instead of working on a single data set, two different data sets were combined and a common data set was studied. Thus, models with general success in classifying arrhythmias were obtained.
More than one arrhythmia can be found in an ECG recording. Since an ECG recording may contain more than one arrhythmia, a threshold value approach was used to classify multi-label ECG recordings. Thus, the trained models were able to detect multiple arrhythmias in ECG recordings.
An ECG recording may contain no arrhythmias. The 'no arrhythmia' class has been defined to classify ECG recordings that do not contain any arrhythmia. Defining the 'no arrhythmia' class is a new approach. By defining the 'no arrhythmia' class, it can be determined whether the ECG recordings contain any arrhythmia. In ECG recordings containing arrhythmia, more than one arrhythmia can be detected with the threshold value approach.
We trained SE-ResNet34 and FCN artificial neural networks to classify arrhythmias detected through ECG recordings. The Squeeze and Excitation (SE) layer enable the network to perform dynamic channel-wise feature recalibration. One-dimensional convolutional network was used for feature extraction from 12-lead ECG recordings in the dataset. The convolutional network used is 34-layer ResNet.
By using the weight function, less weight was given to arrhythmias that occurred more frequently in the data set, and more weight was given to arrhythmias that occurred less frequently. The weight function is given as a parameter while training the model. The studies were conducted for 5, 10, and 15 classes of arrhythmia entities. Training the model on 5 arrhythmias takes less time than on 10 and 15 arrhythmias, since it contains fewer neurons in terms of running time.
Changing the threshold value greatly affects the success of the model. While many arrhythmia classes occur at low threshold values, only a single arrhythmia class occurs at high threshold values. The reason for this is that if no arrhythmia exceeds the threshold value, the arrhythmia with the highest prediction score is considered as the output of the model.
In the FCN model, when working with 5, 10 or 15 arrhythmias, the best results were always obtained when the threshold value was 55%. In the ResNet model, the best results were obtained at 35% threshold values when working with 5 arrhythmias, and at 10% threshold values when working with 10 and 15 arrhythmias.
In the thesis study, it was seen that the FCN model was more successful in detecting arrhythmias than the ResNet model.
In the models created, arrhythmias can be detected with success rates ranging from 60 percent to 90 percent. The current study may help cardiologists make a diagnosis by preventing misinterpretation of ECG signals.
