Detecting a broader spectrum of cardiac pathologies in electrocardiogram data by applying a deep neural network designed to detect a specific cardiac disease. (14th October 2021)

• Record Type: Journal Article
• Authors: Wegener, S; Schmidt, T; Prim, J; Gumpfer, N; Gruen, D; Hannig, J; Guckert, M; Keller, T
• Is Part Of: European heart journal. Volume 42(2021)Supplement 1
• Publication Date: 2021-10-14
• Subjects: Artificial Intelligence (Machine Learning, Deep Learning); Cardiology -- Periodicals; Heart -- Diseases -- Periodicals; 616.12005
• DOI: 10.1093/eurheartj/ehab724.3063
• Languages: English
• ISSNs: 0195-668X

Abstract

Abstract: Background/Introduction: The electrocardiogram (ECG) is a widely used and inexpensive tool that provides extensive insights into the cardiac structure and function. Artificial intelligence (AI) algorithms, especially deep learning (DL) models, are efficient computer based instruments with which large and complex datasets can be processed for identification of e.g. specific diseases. PhysioNet is a NIH research resource for complex signals including a large amount of labelled ECG time-series data. Our aim was to evaluate the diagnostic performance of an AI architecture developed to detect a specific cardiac pathology in a large ECG data set including a broad range of cardiac abnormalities. Methods: The PhysioNet ECG dataset provided as part of the PhysioNet Challenge 2020 consists of five distinct databases with a total of 43100 12-Lead ECG recordings of varying length stemming from patients from China, Russia, Europe and the United States. Each ECG recording is annotated with diagnoses based on a set of 111 possible labels, which express either a cardiac pathology, e.g. atrial flutter or anterior wall ischemia, or unspecific changes in the ECG, e.g. a prolonged qt interval or low qrs voltages. Based on these labels we defined 10 groups merging PhysioNet labels describing related cardiac abnormalities (see Table 1). We adapted a recently published DL model which used raw ECG time-series data of all 12-leads rather than extracted features as model input. This DL model was adapted to the larger number of output variables and then trained on 80% (n=34480 ECGs) of the PhysioNet dataset. The remaining 20% (n=8620 ECGs) of the PhysioNet dataset were used to evaluate the diagnostic performance of the AI model. Sensitivities, specificities and the areas under the receiver operator characteristic curves (AUROC) were used as performance metrices. Results: The AI model, that was initially designed to detect a specific cardiac pathology, performed well in the large PhysioNet dataset providing AUROCs ranging from 0.78 to 0.95 to detect the defined 10 cardiac abnormality groups. Interestingly, the AI model was able to detect disease groups with changes in the chronological sequence of the ECG, e.g. arrhythmia, with comparable precision as disease groups associated primarily with changes in the ECG amplitude like e.g. ischemia. Detailed results are presented in Table 2. Conclusion(s): Our evaluation shows that an AI model that uses raw ECG time-series data rather than extracted features as model input can be easily transferred to other large datasets with different prediction variables. This might also serve as a proof of concept that raw data instead of pre-selected features should be used as model input if developing AI applications for medical use cases. FUNDunding Acknowledgement: Type of funding sources: Public Institution(s). Main funding source(s): FlexiFunds by Forschungscampus Mittelhessen