Mirko Hu — Department of Medicine and Surgery, TecMed Lab, University of Parma #
ECG Signals Revisited with Network Science derived Features #
Since the introduction of the electrocardiogram (ECG) by string galvanometer in 1901 by Prof. Einthoven, the interpretation of cardiac health has remained largely unchanged. Traditionally, experts analyze ECGs for diagnosis, but the increasing volume of data and advancements in computer-based methods necessitate new approaches for feature extraction. Network science is becoming a common language to describe complex systems. In network science, complex systems are represented as nodes connected with edges. Time series can be considered complex systems too and they can be translated into networks by using visibility graphs. It is possible to characterize these complex systems by calculating the graph properties and by using a cartography-based method. Cardiovascular signals, exhibiting strong pseudoperiodic behavior, present challenges in early detection of arrhythmias, a prevalent cardiovascular disorder. Early recognition as well a prediction of arrhythmias can potentially save many lives. We used the 2017 PhysioNet/CinC Challenge database, and we analysed 1,000 short ECG recordings (500 normal and 500 arrhythmic). One approach to ECG analysis via network science involves segmenting signals into chunks and transforming them into visibility graphs. The visibility condition states that it is possible to connect two time points if they are visible to each other, i.e. it is possible to connect the values of two time points without intersecting the values of the time points between them. Then, the multiple graphs from one ECG were overlapped, and the weights were normalised to obtain a weighted graph with weights between 0 and 1. We used an arbitrary threshold of 0.50 to cut the noisy edges. We obtained, in this way, a unique representative graph for the ECG of a subject. To extract the features that were used to classify an ECG, we performed community detection through Louvain algorithm, and we identified the roles of each node in the graph. The roles depend on the position of a node inside the community. We also calculated some properties of the graph spanning from the diameter to the density. The percentage of the number of nodes for each role, the total number of nodes, the average degree, the density, the diameter, and the average clustering were used as features for a random forest classifier. After optimising the hyperparameters of the machine learning classifier, we obtained an accuracy of 74% and an AUC of 0.81 on the test set (300 ECG recordings, 150 normal and 150 arrhythmic). This work can pave the path for revisiting the traditional ways of reading ECG based on the analysis of the typical ECG waves, such as the QRS complex, the P and the T waves. This work also presents an innovative way of using a cartography-based analysis of the network and of extracting new numerical features from an ECG signal. Finally, this work can help machines to better recognise the arrhythmic patterns absent in the normal signals.