CardioInceptionNet: A Novel CNN Architecture for ECG Abnormalities Detection

Oles Dobosevych1, Illia Kachko2, Bohdan Petryshak1
1The Machine Learning Lab at UCU, 2SoftServe Research Group, The Machine Learning Lab at UCU


Abstract

The Electrocardiogram(ECG) plays a crucial role in detecting different types of cardiac abnormalities in clinical practice. However, the arrhythmia types variations, high inter-patient variability make this task a challenging issue. There is a wide range of RNN-based approaches, which researchers employ for tasks on ECG signal disease detection. The existing traditional CNN architectures, contrary to RNN approaches, do not take into account the temporal aspect of the ECG signals. Moreover, they have a small size of the convolutional kernels, which leads to the hindering of the performance, especially for the AF, PAC, and PVC pathologies, which require huge receptive field size of the network to be detected.

      In this work, the cardiac abnormalities detection task from the ECG signal was treated as a time series classification problem(TSC). We propose a CardioInceptionNet, an adopted for signal data 1D convolutional neural network, trained to classify eight types of cardiac abnormalities. Inspired by the Inception-v4 architecture, CardioInceptionNet consists of 3 inception modules. The first key component of each module is the "bottleneck" layer, which significantly reduces the dimensionality of the data and model's complexity, mitigating overfitting problems. The second major component is huge sliding multiple filters simultaneously on the same input signal with length ∈ {10, 150, 300}, which effectively learns long-range abnormality patterns within P-waves, QRS complexes, and cardiac cycles. We also introduce the parallel MaxPooling operation to make our model invariant to small signal perturbations. We utilized the residual connections along with batch normalization layers to improve the convergence of the model during training. 

      We evaluated our algorithm on 688 patients using the Stratified five-Folds Cross-Validation technique and obtained an F_1 accuracy of 75% for classifying. We got the 0.46 F_2 Score, 0.28 G_2 Score, and 0.36 Geometric Mean on the public leaderboard. 

The proposed algorithm is both accurate and lightweight.