Rule-Based methods and Deep Learning Networks for Automatic Classification of ECG

giovanni bortolan1 and Ivaylo Christov2
1IN-CNR, 2Institute of Biophysics and Biomedical Engineering - Bulg. Accad. of Sci


Objective: The objective of the material is to explore the potentiality of combining a classical rule-based method with a Deep Learning method for automatic classification of ECG in the framework of the PhysioNet Chal-lenge 2020. Materials: The training set is a public database collected in China and consists of 6877 (3699 males and 3178 females) 12-leads ECG recordings lasting from 6 seconds to 60 seconds. Each ECG recording has one or more diagnostic labels: Normal sinus rhythm, Atrial Fibrillation (AF), First-degree Atrio Ventricular Block (I-AVB), Premature Atrial and Ventricular Complex (PAC, PVC), Left and Right Bundle Branch Block (LBBB, RBBB), ST-segment depression and elevation (STD, STE). Method: The rule-based method is using morphological and time-frequency ECG descriptors, characterizing each diagnostic labels. These rules have been extracted from the knowledge-base of a physician, with no direct learning procedure in this phase. A refinement in the training database will be tested in the second phase. For example, more specific limits for PVC, PAC, STE or STD can be usefully extracted from the training set. The Deep Learning method consider both raw ECG signals and median beat signals. These data are processed by continuous wavelet transform analysis obtaining a time-frequency domain representation, with the genera-tion of specific images. These images are then used for training Convolution-al Neural Networks for ECG diagnostic classification. Three pre-trained CNN for image classification have been initially used, GoogleNet, SqueezeNet , and Alex Network . Results: Unofficial results of the classification accuracy on an independ-ent ECG testing set, are reported in the Table. These results suggest the possi-bility to increase the accuracy with a more robust learning process for DL and the use of appropriate rule-refinement for benefitting of the training set. In addition, the possibility to explore the potentiality of combining the two methods will be and studied