A Critical Look At Studies Applying Over-Sampling On The Tpehgdb Dataset

Preterm birth is the leading cause of death among young children and has a large prevalence globally. Machine learning models, based on features extracted from clinical sources such as electronic patient files, yield promising results. In this study, we review similar studies that constructed predictive models based on a publicly available dataset, called the Term-Preterm EHG Database (TPEHGDB), which contains electrohysterogram signals on top of clinical data. These studies often report near-perfect prediction results, by applying over-sampling as a means of data augmentation. We reconstruct these results to show that they can only be achieved when data augmentation is applied on the entire dataset prior to partitioning into training and testing set. This results in (i) samples that are highly correlated to data points from the test set are introduced and added to the training set, and (ii) artificial samples that are highly correlated to points from the training set being added to the test set. Many previously reported results therefore carry little meaning in terms of the actual effectiveness of the model in making predictions on unseen data in a real-world setting. After focusing on the danger of applying over-sampling strategies before data partitioning, we present a realistic baseline for the TPEHGDB dataset and show how the predictive performance and clinical use can be improved by incorporating features from electrohysterogram sensors and by applying over-sampling on the training set.