Paper ECG Digitization App

Today, machine learning algorithms are used to study heart disease. To work at their best performance the algorithms mentioned above need a large amount of data to process. Unfortunately, most of the electrocardiograms (ECGs) collected in the archives by cardiologists is still on paper. Therefore, a preparatory work for the use of the citated algorithms is the creation of a dataset composed of images of leads and their conversion to electrical signals. For this purpose, the ECG-D application was created and this thesis explains its architecture, features, achieved results and limitations. To digitize the ECGs, a digitization process was designed and included in application. It is divided in 4 phases. In order of execution they are: preparation, background removal, conversion from image to signal e digital signal display. In first phase, after selecting the image or pdf file, some actions such as rotation, scale factor detection, image resizing and cropping are performed to prepare the image for background removal, second phase, where the goal is to get an image without grid. This is the most important stage of the whole process, if the grid is removed correctly the final result will be good. To remove the grid and noise from the image there are two methods. One is based on the histogram and the other one on morphology. The histogram method is used when the grid is colored or has a different grayscale of the signal. Instead the Morphology method is used when the grid has the same color as the signal. Then, in converting the image into electrical signals phase, the pixel position in the image is converted into an electrical signal. To check the conversion results, the application includes two tools: • The first is a line graph showing the converted electrical signal; • The second is an image of the electrical signal overlapped to the original clipping. In addition, an advanced function that allows you to correct the peak values has been developed to mitigate error introduced by cardiograph during the drawing process. Also editing image function has been included in the application to retouch the image when it is damaged or leads in it are overlapped. Application has been validated measuring correlation between real value obtained from electrocardiograph and signal digitized using ECG-D application. Result obtained during validation phase has been at least 99.5% using a set of 40 ECGs. To conclude, we need take in account that most of the cardiographs present on the market from many years are digital, therefore from them, in addition to the paper ECG, the digitized signal is also obtained. Despite this, the number of clinically “valuable” ECGs that are still stored in paper records and that are often retrieved to derive measurements is still very large and justifies the need for reliable tools to digitize these leads and archive it in digital format into database usable for current and future research. Therefore, the availability of tools such as ECG-D should not, in any way, encourage paper collection, because, the experience learned from the validation study showed that some information is already lost with the electrocardiograph printing process and even with 100% accurate reconstruction, we will never be able to fully restore the original information. Therefore, the use of tools such as ECG-D can be limited only to situations where digital ECGs are no longer available or in retrospective studies with the aim of building representative digital databases.