Improved convolutional neural network with feature selection for imbalanced ECG Multi-Factor classification

•Add a differential sequence to reduce the impact of noise in ECG signal.•Add a variable nonlinear layer to enhance the non-linear capability of the model.•Adaptive Learning rate is used to converge the training process.•The performance of proposed CNN is better than the other widely used algorithms.•The proposed algorithm can serve as an adjunct tool to assist clinicians in confirming their diagnosis. In this paper, an improved convolutional neural network (CNN) named Feature Selection CNN (FS-CNN) and Synthetic Minority Over-sampling Technique (SMOTE) named Multiply Stochastic SMOTE (MS-SMOTE) are proposed for imbalanced ECG multi-factor classification. FS-CNN includes convolutional expression layer, multi-residual block, and decision layer. The primary purpose of FS-CNN is to find the most influential features during automatic model optimization dynamically. MS-SMOTE combines the advantage of SMOTE and borderline SMOTE to balance the number of different categories. By dynamically selecting SMOTE and borderline SMOTE based on the proportion of data, the synthetic data can have a better category property. Comparing to other wide-used algorithms, our method is better than most other algorithms in many indicators. And in simulation 2, our method can deal with a large number of ECG multi-factor data classifications. Also, two self-controlled experiments are designed to examine how different parameters affect the result of the above problem.