A Machine Learning System for Automatic Detection of Preterm Activity Using Artificial Neural Networks and Uterine Electromyography Data

Background: Preterm births are babies that are born before 37 weeks of gestation. The premature delivery of babies is regarded as a major global public health issue with those affected at greater risk of developing short and long-term complications. The care provided for premature infants has significantly improved. However, it has had no impact on reducing the prevalence of preterm birth. Therefore, a better understanding of why preterm births occur is needed. Methods: Electromyography is used to capture electrical activity in the uterus to help treat and understand the condition, which is time consuming and expensive. This has led to a recent interest in automated detection of the electromyography correlates of preterm activity. This paper explores this idea further using artificial neural networks to classify term and preterm records, using an open dataset containing 300 records of uterine electromyography signals. The Synthetic Minority Oversampling TEchnique is used to oversample the minority preterm class (38 records) to address the issues found in unbalanced datasets and classification. Results: Our approach shows an improvement on existing studies with 94.56% for sensitivity, 87.83% for specificity, and 94% for the area under the curve with 9% global error when using the Multilayer perceptron neural network trained using the Levenberg-Marquardt algorithm. Discussion: The Multilayer perceptron neural network trained using the Levenberg-Marquardt algorithm produced the best results, which is trained using Newton’s method of least squares optimization and is an efficient learning algorithm for neural networks that have a few hundred weights, despite being computationally expensive.