Comparative Analysis of Data Augmentation Approaches for Blood Pressure Prediction.

Area of Science:

• Biomedical Engineering
• Data Science
• Time-Series Analysis

Background:

• Blood pressure (BP) time-series data exhibit non-stationarities and irregularities.
• Accurate BP forecasting is crucial for patient monitoring and management.
• Existing data augmentation techniques may not fully address the complexities of BP data.

Purpose of the Study:

• To explore and compare traditional and advanced data augmentation techniques for BP time-series data.
• To evaluate the impact of these augmentation methods on forecasting model performance.
• To identify optimal augmentation strategies for improving BP time-series analysis.

Main Methods:

• Application of traditional augmentation methods: windowing, magnitude wrapping, and cropping.
• Investigation of Generative Adversarial Networks (GANs) for advanced data augmentation.
• Forecasting using a Jump Neural Network model with varying input window widths and prediction horizons.
• Analysis on a heterogeneous patient population dataset.

Main Results:

• Traditional methods effectively handle BP time-series non-stationarities.
• Generative Adversarial Networks (GANs) demonstrate potential for enhancing forecasting accuracy.
• Different augmentation techniques exhibit subtle but significant effects on prediction outcomes.
• Performance varied based on input window width and prediction horizon.

Conclusions:

• Data augmentation significantly impacts BP time-series forecasting accuracy.
• Generative Adversarial Networks (GANs) offer a promising avenue for improving BP prediction models.
• The choice of augmentation method should consider the specific characteristics of the BP time-series data and the forecasting task.