A Novel Technique in Classifying Heart Diseases based on Electrocardiogram (ECG) Signals using Deep Learning and Spectrogram Image Analysis

Deep learning is among the most reliable approaches in analyzing physiological signals for healthcare applications. This makes data mining strategies more useful and reliable in analyzing a large group of data like medical records. Through this technology breakthrough, the authors established a novel approach in examining electrocardiogram (ECG) records and classify them on what kind of heart disease based from its pattern using spectrogram image analysis. These records underwent data training as part of the deep learning process. Data trained are in the form of extracted ECG logs in comma-separated values (CSV) format from different patient cases. These databases are downloaded from Physionet databases, which offers free access to an extensive collection of recorded physiological data. Once done downloading, these databases were categorized per disease type. Records of diseases that were in the same category must be saved in the same folder. These CSV files passed through a preprocessing procedure through filtering CSV log files and cut each ECG reading into 10-second duration and save it as text file. These text files were converted into their spectrogram images equivalent. The converted text files to spectrogram images were still categorized per disease type. The file name of each folder containing the spectrogram images for each disease type served as the label map for the output classifier. The final layer was retrained for dataset development through the use of Google's Inception V3 model that serves as a base learning model. Fine tuning was established through setting its training rate to 0.001, training batch size to 500, a validation batch size of 200 and a total number of training steps of 500,000. These settings had achieved the optimum result of 100% validation and training accuracy each. The result of final test accuracy is 98.9% and had generated a frozen model for the classifier. Its accuracy can improve if there are more available and reliable ECG records. Using the frozen model, the same procedures were done in preparing the dataset to be used for the testing stage. The result of the classification of disease is shown through the percentage value of how likely is the input spectrogram image based from disease categories. The authors evaluated the proposed technique and obtained a significant performance on the classification of types of heart diseases based on available databases from Physionet.