Comparative Analysis of Data Augmentation Approaches for Blood Pressure Prediction

This study explores several data augmentation techniques designed especially for Blood Pressure (BP) time-series data. First, we discuss traditional methods like windowing, magnitude wrapping and cropping, which work especially well for handling the natural non-stationarities and irregularities of BP time series. Next, we investigate the use of Generative Adversarial Networks (GANs) to determine whether more advanced data augmentation methods can lead to better forecasting model performance. We apply these methods to forecasting using different input window widths and prediction horizons on a heterogeneous patient population. Our analysis is based on the application of a Jump Neural Network model, which was selected due to its flexibility and effectiveness in processing time-series data with different properties. Our findings shed some light on the subtle effects of each augmentation method on forecasting accuracy and suggest the best methods for improving BP time-series analysis.