Fetal Movement Detection by Wearable Accelerometer Duo Based on Machine Learning
Monitoring fetal movement (FetMov) is crucial during pregnancy as it indicates fetus health. The majority of women experiencing stillbirth experienced reduced fetal movement, which pregnant women felt before the diagnosis. FetMov may contain important information about biological indexes for fetal well-being estimation.
FetMov is the result of the central nervous and musculoskeletal systems. A decrease in FetMov may imply fetal growth restriction or a nuchal cord. Monitoring the FetMov and gathering helpful information about the fetus's physical or physiological development status, even in out-of-hospital settings, is crucial. Thus, an accurate FetMov monitoring device is required to monitor fetuses and mothers-to-be.
The researchers developed wearable techniques to monitor FetMov with performance comparable to the ultrasound technique. A passive and wearable device with two accelerometers was designed to sense subtle motion in the abdomen of pregnant women. Clinical trials with ultrasound as the gold standard and the wearable device on 20 pregnant women demonstrated the system’s capability. Compared with ultrasound, the system could classify FetMov with an accuracy of 86.6% and F1 of 84.2%, which is much better than maternal perception.
Two mCube MC3672 were chosen as the accelerometers, and its acquired micromotion data on the abdomen was transmitted to Nordic nRF52840. The study was carried out on 20 pregnant women with a mean age of 28.1 years (SD 2.8 years) and a mean fetal development time of 38.4 weeks (SD 3.2 weeks). A portable ultrasound machine, GE LOGIQ V2, was used for FetMov detection. Due to the environmental factors and signal drifting of the sensors during the experiment, the raw data was preprocessed.
After analyzing the FetMov waveform from several features including statistical, morphological, and wavelet, binary classification frameworks were created to recognize FetMov. The Synthetic Minority Oversampling Technique (SMOTE) function was used to increase the number of minority cases to make the distribution of the two data types more balanced before data training. Nine machine learning techniques in Scikit-learn were compared side by side to recognize FetMov automatically.
The performance of maternal perception has an F1 score of 12%, and variation between pregnant women was clearly observed. Compared with the ultrasound as the gold standard, the high-performance metrics of Extra Trees Classifier with an accuracy of 86.6%, recall of 82.4%, precision of 86.1%, and F1 of 84.2% indicate that it could potentially replace the active ultrasound method for accurate FetMov monitoring.
The researchers designed and developed a wearable device based on two acceleration sensors and compared the metrics of nine machine-learning methods with ultrasound as the gold standard. The device designed can be worn on the abdomen of pregnant women, and has low energy consumption, thus making it suitable for long-term. It can realize accurate detection and continuous monitoring of FetMov, an important index to evaluate the health status of a fetus in the mother's womb. However, the limitation of the model is that more training data is needed. The research is still ongoing.