Non-linear signal analysis of uterine electromiogrammes for predicting pre-term labour

Gašper Fele Žorž (2010) Non-linear signal analysis of uterine electromiogrammes for predicting pre-term labour. PhD thesis.

PDF Download (6Mb)

Abstract

Our aim was to analyze uterine electrical activity using various signal processing techniques. We carried out our research on 1211 records of uterine electrical activity which were recorded between the years 1997 and 2005. Most of the records contained three channels to record the electrical activity. 53 records also contained an estimate of the intra-uterine pressure as measured by a tocograph. Most of the records were taken either around the 22nd or around the 31st week of gestation. The duration of most records was 30 minutes. In addition, for most records accompanying data was collected. We checked the records for excessive noise and rated them accordingly. The accompanying information was collected from multiple sources and organized. We stored the records in a relational database. The database contained 760 records where all three signals were almost certainly good, and for which all neccessary accompanying data was available, recorded during 532 pregnancies. We implemented a system for the storage of records, their signals, their accompanying data and the results of various computations, performed on the signals. We filtered the signals using six band-pass filters. The lower band-pass frequency limits were 0.08 Hz and 0.3 Hz. The upper band-pass frequency limits were 2.5 Hz, 3 Hz and 4 Hz. We used Butterworth filters with a double-pass filtering scheme. This scheme maintains most benefits of infinite impulse response filters while eliminating the group delay which might be problematic. We implemented a graphical user interface to view the records using various signal visualization techniques. The user interface supported the display of signals as time series and their power spectra. In addition, the interface could display spectrogram and the Wigner-Ville time-frequency distribution of each signal. The results of various calculations were also displayed for each signal. The records and the signals to be displayed were selected by writing simple SQL queries into entry fields in the user interface. While observing the records which contained an estimate of intra-uterine pressure, we observed changes in the electrical activity of the uterus during contractions. The changes were especially evident while observing the Wigner-Ville time-frequency distribution of the signals. We also calculated the short-time cross-correlation coefficients between the signals of these records. We observed that the correlation between signals was almost always highest during the same time instant. The peaks of the cross-correlation coefficients usually rose during contractions. In addition to visualising of records, we also calculated some properties of the signals using different signal processing techniques. We divided the techniques into two groups - the linear and non-linear techniques. The linear techniques were based on the power spectrum of each signal. They included the root mean square value of the signal, the peak frequency of the power spectrum, it's median frequency and the first zero-crossing of the autocorellation coefficients. The non-linear techniques were based on the estimation of complexity of each signal. They included the maximum Lyapunov exponent and the correlation dimension, both of which are based on the reconstructed phase-space of the system, and sample entropy along with it's extension, the multi-scale sample entropy, which are based on the self-similarity of each signal. The non-linear signal processing techniques used are computationally more demanding than the linear ones. The time complexity of the most widely used algorithm for the calculation of sample entropy is O(N2), where N represents the length of the signal. We developed a new algorithm to calculate the sample entropy of a signal. The time-complexity of the new algorithm on typical signals is O(N log(N)). As part of the implementation of the new algorithm, we also created a fast implementation of the skip-list data-structure. We divided the records where all signals with their accompanying data were good and where labor was spontaneous, into multiple groups. We formed the groups according to the duration of pregnancy and the time of recording. This yielded four groups of records - those records where birth was premature, those records where birth was on term; those records which were taken before the 26th week of gestation and those records which were taken during or after the 26th week of gestation. We then used the Student's t-test to calculate the probabilities that the means of various signal properties were the same across pairs of groups. We thus obtained six probabilities, which we then used to identify the most promissing techniques. We observed that the average value for some properties of records, taken before the 26th week of gestation, was different when calculated for those records where birth was premature versus those records where birth was on term. On the basis of the Student's t-test we concluded that the most promising signal processing techniques were the median frequency of the power spectrum and the sample entropy. In cases where two records were available for each pregnancy, we used the Student's t-test for paired samples to estimate the changes of various signal properties throughout the pregnancy. We also calculated the probability that the average changes were different for those records where birth was premature than for those records where birth was on term. We also tried to classify records on the basis of the calculated properties. We attempted the classification in pairs. We first tried to classify those records where birth was premature versus those where birth was on term. Then, we tried to classify those records which were taken before the 26th week of gestation versus those which were taken later. We repeated the classification of pre-term versus term records for those records which were taken before the 26th week of gestation and then for those taken later. We also repeated the classification of records taken before the 26th week of gestation versus those taken later, among those records where birth was premature and among those records where birth was on term. For classification we used the naive Bayesian classifier and decision trees. We tested each classifier using three tests - first by testing the classifier on the learning set, then by using cross-validation and finally using the leave-one-out approach. The results of testing the classifiers on the learning sets were excessively optimistic. Decision trees in particular were shown to be prone to overfitting. The Naive Bayesian classifier was better in this regard and generally performed better, despite it's simplicity. In general, the results of classification were worse than expected. We also tried to classify records by using features, obtained using principal components analysis of their original features. The classification results using principal components analysis were even worse than the results from classifying the records using their original features. In particular, the performance of decision trees when using principal components analysis often turned out to be worse than randomly guessing the classes of the records.

Actions (login required)