Non-invasive fetal electrocardiogram classification using spatial correlation

  • M.J.M. Hermans

Student thesis: Master


In Europe and many developed countries the preterm birth rate is generally 5-9 %, and in the USA it has even risen to 12-13 % in the last decades. Preterm birth is the major cause of neonatal mortality in developed countries. Premature infants are at greater risk for short and long term complications, including disabilities and impediments in growth and mental development. The cause for premature birth is in many situations difficult to determine; many factors appear to be associated with the occurrence of premature birth, making the prevention of premature birth a challenging proposition. Fetal monitoring can support the physician in making the often vitally important decision whether or not to induce labor artificially. Nowadays, the most common used technique for fetal monitoring is cardiotocography (CTG). However, information from Doppler CTG alone is not sufficient for making important decisions regarding the treatment of the fetus and the mother in case of fetal distress. The predictive value of the CTG can be increased when it is combined with ST analysis of the electrocardiogram (ECG) of the fetus. The STAN® device (Neoventa, Sweden) uses this combination by attaching a single spiral electrode to the fetal scalp to obtain the fetal ECG (fECG) and to perform ST analysis. The major drawback of the STAN® device is, however, its invasiveness. It can only be performed when the fetal membranes are ruptured and sufficient cervical dilatation is present. In order to get a better insight into the fetal condition during pregnancy, it would be advantageous to develop a method for fetal monitoring that operates non-invasively, and provides additional information with respect to the CTG. In collaboration with the Eindhoven University of Technology, at the Máxima Medisch Centrum in Veldhoven, The Netherlands, researchers have developed an algorithm to obtain the fECG from the abdominal recordings for singleton pregnancies. The algorithm is capable of estimating the maternal ECG (mECG) and subtracting it from the abdominal recordings, but due to the low signal to noise ratio (SNR), the signals remaining after subtraction of the mECG are often not suitable for extracting the fECG. Additionally, at the moment, the calculation time of the algorithm is too long to permit online fetal heart rate detection. A solution to decrease the calculation time significantly could be by excluding several abdominal signals for further analysis. The main goal of this project was therefore to reduce the calculation time of the fECG extraction by excluding a number of abdominal signals from further analysis. In order to achieve this, the abdominal signals need to be screened in order to assess their quality. In this report, a technique is presented that classifies the abdominal recordings based on their spatial correlation. The technique is referred to as the classification of fetal ECG signals using spatial correlation (CLASC). CLASC considers a priori knowledge on the abdominal electrode configuration and fetal heart activity. It obtains a two dimensional representation of the fetal vectorcardiogram (fVCG) by using a patient-tailored vectorcardiography approach to spatially combine the abdominal fECG signals. However, due to the low SNR of the abdominal fECG recordings, this fVCG representation is significantly corrupted by noise. However, in spite of the low SNR, the fVCG can be Once used to locate the positions of the fetal R-peaks and to define individual fECG complexes. These ECG complexes can, in turn, be enhanced by averaging the detected ECG complexes, synchronized on their R-peaks. The resulting average ECG complexes can be used to obtain an fVCG, again using the patient-tailored vectorcardiography approach mentioned before, with increased signal to noise ratio. This fVCG, since it originates from averaged ECG complexes, is referred to as the average VCG (AVCG). By projecting the AVCG onto the lead vectors that indicate the electrode positions, and comparing the resulting projected fECG signals to the originally determined average fECG complex, the fECG signal exhibiting the lowest correlation can be identified. This signal is subsequently considered to contain the least useful, mutual fECG information and is therefore classified in the last position and omitted from further processing. From the remaining average ECG complexes, the process (i.e. calculation of the AVCG, projecting this AVCG onto the lead vectors, and omitting the signal with lowest correlation) described above is repeated until full classification of the abdominal recording is achieved by CLASC. By limiting further processing and analysis of the abdominal fECG signals to the best (e.g. four) signals, calculation times can be decreased significantly, while only a limited amount of fECG information is lost. In order to assess the performance of the fetal R-peak detection of CLASC, it is compared to a widely spread technique called principal component analysis (PCA). In general, CLASC significantly outperforms PCA on abdominal recordings with a large SNR, but is beaten by PCA on low SNR recordings, albeit marginally. The classification part of CLASC is compared to an average visual selection (AVS) which is an average of the visual selection by multiple researchers in the MMC. In general, when both techniques define the best four fECG signals, at least 3 out of four are the same. The advantage of using CLASC is that it operates fast, automatically, and always produces the same results, in contrast with the visual selection, which is dependent of the observer. So as a general conclusion, although CLASC should be enhanced in the future, it already performs quite accurate and fast, and takes a huge step towards future online beat-to-beat heart rate detection and fECG extraction from abdominal recordings.
Date of Award31 Jan 2010
Original languageEnglish
SupervisorJan W.M. Bergmans (Supervisor 1), Richard A.H. Engeln (Supervisor 2), K. Kopinga (Supervisor 2), J.O.E.H. van Laar (Supervisor 2), Rik Vullings (Supervisor 2) & P.F.F. Wijn (Supervisor 2)

Cite this