A Comparative Analysis of Time Series Data Augmentation Methods in the Identification of Diabetic Neuropathies Based on Deep Learning Algorithms

Non-invasive and reliable methods are essential in the diagnostics and treatment planning of diabetic neuropathies. Forecasting models based on postural data seem to be a promising solution to this problem. However, the performance of machine learning models is often hindered by limited number of observations and imbalanced datasets. This research work focuses on an empirical comparative analysis of data augmentation techniques applied to time series analysis, in the domain of diabetic neuropathy detection. Building upon a preprocessed dataset, a suite of data augmentation techniques tailored to time series data are evaluated. Multilayer perceptrons and convolutional neural networks were trained using augmented datasets. Two strategies were employed for training and validation. Model performance was evaluated based on the ability to generalize from augmented to real-world data. These results suggest that data augmentation can be a feasible and reliable approach for prediction diabetic neuropathies based on postural time series data.