Fetal Movement Detection using Long Short-Term Memory Network

The use of fetal movements to estimate fetal health is the subject of focus and is preferred over conventional methods of monitoring fetal health. In this paper, the possibility of using data from an accelerometer attached to the abdomen of the mother to detect fetal movements using recurrent neural networks in realtime is analyzed. Long Short-Term Memory (LSTM) layers were incorporated into the network to ensure that useful features are remembered over multiple data points. Data preprocessing was carried out to ensure proper training of the network. Acceptable results were obtained and the network was trained well with the given data set. Initially, the raw accelerometer readings were used as the input to the LSTM network and since the raw data did not yield good results, the LSTM network was modified to use spectrogram of the accelerometer data as the input which made a dramatic improvement on the results obtained. The neural networks implemented in this paper had fewer layers mainly due to the absence of enough data because a huge amount of data is needed for a deep network. Furthermore, with a larger data set, a deeper network can be trained which will yield more reliable results which would ultimately give an algorithm that can detect fetal movement accurately. This paper hence demonstrates the fact that the use of recurrent neural networks to identify fetal movement is possible and can yield satisfactory results.