Performance assessment of AI tools for digitizing ECG scans

Abstract

The digitization of electrocardiogram (ECG) signals recorded on paper is a very challenging task usually prone to errors and inaccuracies. Until now there is no available tool which can perform that task both universally and in a fully automated way. ECG is a well established medical modality to record the activity of the human heart which dates to over 100 years ago. Medical experts can diagnose potential heart irregularities by interpreting the recorded signals on ECG papers. There are millions of ECGs worldwide and the digitization of them is of paramount importance for research, analysis and diagnosis in medicine. By recognizing patterns in the digitized ECGs, artificial intelligence algorithms can predict cardiovascular diseases and help clinicians to make better medical diagnoses. Therefore, the development or the improvement of an automated tool that will be able to digitize massively ECG signals at once, is essential. In this thesis we describe and develop a tool for digitizing ECGs and we test its performance using ECG scans provided by Akershus University Hospital. In particular, we introduce some improvements which can make it operate in a more automated way. The original tool has several parameters in its various steps that prevent it from being fully automated. However, with proper further improvements, it has a great potential to become fully automated at least for ECG scans similar to the ones in the database of Akershus University Hospital. The current master thesis was held during the last semester of the ACIT’s master program of Oslo Metropolitan University, namely from January 1st until May 15th.