Spectrogram and Deep Neural Network Analysis in Detecting Paroxysmal Atrial Fibrillation with Bottleneck Layers and Cross Entropy Approach

Paroxysmal AF (PAF) is a form of atrial fibrillation (AF) that is generally clinically silent and undetected. AF is a type of heart disease called cardiac arrhythmia. Automatic detection of AF could make a significant contribution to early diagnosis, control and prevention of chronic AF complications. In this paper, authors presented a novel algorithm through spectrogram and deep learning neural network analysis in detecting paroxysmal AF from image data segments. This method does not require the detection of P and/or R peaks which is a preprocessing step required by many existing algorithms. The PAF Prediction Challenge Database from Physionet.org were used as learning set which composed of 50 record sets. These records were converted into 7,000 PAF and 964 healthy data segments. Each data segment has 5 mins-duration and converted it to graph images. These graph images are then converted into spectrogram to visualize the frequency band present in the spectrum. In this process, ECG numerical values were interpreted into spectrogram form. Spectrogram images are cropped to remove unnecessary markings from the graphing and spectrogram processes. Cropped spectrogram images are then grouped into separate folders according to type. The produced datasets are then fed into training using 500,000 training steps. The algorithm is integrated with TensorFlow CPU version 1.5 and Inception V3 model to take advantage of its astonishing way on how it analyzes images. The deep learning neural network involves a bottleneck layer which uses lesser neurons to reduce the number of feature maps in the network to get the best loss during training. In order to have a faster learning rate, the cross-entropy cost function was used. The final accuracy test from the training reached as high as 96.8%. An actual test for identified PAF and healthy datasets from Physionet.org were performed and all are correctly predicted and thus could be able to classify other different diseases based from converted ECG numerical values. Furthermore, this paper established a low-powered workstation’s requirement for implementation because it only requires at least a dual core processor and 2 GB of RAM.