Abstract:
Fetal activity is an important indicator of a healthy fetus and its normal growth in the womb as they are the reliable sign of normal functional central nervous and musculoskeletal system of the fetus. Electronic fetal monitors are commonly used worldwide to monitor fetal health. Due to the development in electronics, medical and computer technologies, fetal monitors are now improved in functionality and smaller in size, but the use of these monitors is still restricted to hospital premises. This research work focuses on the efficient classification of fetal movement from the normal human activity of the mother. Tri-axial accelerometers and acoustic sensors are commonly used in previous works to detect fetal movement and most of the techniques in different available research are restricted to their use in hospitals and clinics. Two different datasets of fetal movement are used in this research work. The datasets contain measurements from single accelerometer for fetal movement detection. After collection and pre-processing of the 3D accelerometer measurements dataset to detect fetal movement from multiple pregnant women, different state of the art machine learning algorithms are implemented to classify the fetal movement from maternal body movement with relative degree of accuracy. Among all employed algorithms, the Extreme Gradient Boost algorithm demonstrates superior performance in classifying fetal movement on Mendeley and Zenodo fetal movement dataset, achieving an accuracy of 94.58% and 96.01%, respectively. Moreover, we also developed multi-label classification algorithms for the classification of fetal movement, laugh, and other maternal body movements. Extra Trees classifier shows 93.72 % accuracy in multi-label classification. Furthermore, the accuracy of deep learning algorithms for multi-label classification of fetal movement, laugh, and other maternal body movements is presented in this research work. The convolutional neural network with 1D convolutional layer shows the high test accuracy of 88.96 % with a test loss of 18.67 %. The efficiency of different classifiers is tested on real time data collected in an unconstraint environment from MyNeuroHealth application. As expected, graph and tree based classifiers provided the highest accuracy on spatially and temporally correlated accelerometry data which was collected from 13 pregnant human subjects.
