Nonlinear Temporal Analysis of Uterine EMG for Preterm Birth Classification

Premature birth, one of the root causes of maternal mortality and childhood morbidity, is growing at an increased rate worldwide. Several screening tests used in clinical settings to predict preterm births do not show satisfactory results. Electrohysterography, a noninvasive automated method of monitoring uterine contractions during labor, has shown to be effective in the classification task. In this work, we explored the potential use of several temporal nonlinear parameters for carrying out classification process. A publicly available TPEHG DB (Term Preterm Electrohysterogram Database) containing 262 term records and 38 preterm records was used to conduct the study. To execute effective classification tasks, Synthetic Minority Over-Sampling Technique (SMOTE) was applied on unbalanced dataset to generate equal number of training dataset by oversampling the minority class. New suitable nonlinear features have been addresed in this study other than the features used in the previous works to compare the performance of the classifiers. The analysis showed that extremely randomized trees or extra trees classifier shows significant improvement by using these features in the classification process of term and preterm records with cross validation accuracy 91.4%, specificity 90.2% and sensitivity 95.5%. The proposed approach could be combined with other methods to excel in the existing classification performance.