Generating Healthcare Time Series Data for Improving Diagnostic Accuracy of Deep Neural Networks

Data scarcity and class imbalance are common occurrences in healthcare datasets and have an adverse effect on classification performance of machine learning models. Artificial data generation in various applications can be used to handle these challenges. This article proposes a guided evolutionary synthesizer (GES), a tool derived from principles of genetic algorithm, and designed to generate artificial healthcare time series data for improving classification performance of machine learning models. We conducted a series of promising and confirmatory preliminary experiments performance using traditional machine learning and nonresidual convolutional neural network models to evaluate the effectiveness of GES synthetic on data classification. Motivated by the preliminary results, we conducted eight detailed experiments using residual neural network (ResNet), which demonstrated the flexibility of GES and the effectiveness of GES synthetic data in improving the classification performance of deep neural networks. These experiments use GES generated electrocardiogram (ECG) and electroencephalogram (EEG) datasets. Our findings show that models trained with GES synthetic data performed better than models trained with regular perturbed data, had better diagnostic performance for both EEG and ECG datasets, achieved better performance with lower training volume for EEG dataset, eased learning biases seen in the literature for ECG normal sinus and premature ventricular complex rhythms, and had better classification outcomes than comparable related work models trained with similar ECG classes.