* Asst Professor E. Salamalekis, 30 Roumelis Street, Halandri, 15233 Athens, Greece.
Objective To develop a computerised system that will assist the early diagnosis of fetal hypoxia and to investigate the relationship between the fetal heart rate variability and the fetal pulse oximetry recordings.
Design Retrospective off-line analysis of cardiotocogram and FSpO2 recordings.
Setting The Maternity Unit of the 2nd Department of Obstetrics and Gynaecology, Aretaieion Hospital, University of Athens.
Population Sixty-one women of more than 37 weeks of gestation were monitored throughout labour.
Methods Multiresolution wavelet analysis was applied in each 10-minute period of second stage of labour focussing on long term variability changes in different frequency ranges and statistical analysis was performed in the associated 10-minute FSpO2 recordings. Self-organising map neural network was used to categorise the different 10-minute fetal heart rate patterns and the associated 10-minute FSpO2 recordings.
Main outcome measures Umbilical artery pH of ≤7.20 and Apgar score at 5 minutes of ≤7 formed the inclusion criteria of the risk group.
Results After using k-means clustering algorithm, the two-dimensional output layer of the self-organising map neural network was divided into three distinct clusters. All the cases that mapped in cluster 3 belonged in the risk group except one. The sensitivity of the system was 83.3% and the specificity 97.9% for the detection of risk group cases.
Conclusions A relationship between the fetal heart rate variability in different frequency ranges and the time in which FSpO2 is less than 30% was noticed. Fetal pulse oximetry seems to be an important additional source of information. Computerised analysis of the fetal heart rate monitoring and pulse oximetry recordings is a promising technique in objective intrapartum diagnosis of fetal hypoxia. Further evaluation of this technique is mandatory to evaluate its efficacy and reliability in interpreting fetal heart rate recordings.
During the last decades, fetal heart rate monitoring has been widely used for ante- and intrapartum monitoring and assessment of fetal wellbeing. It is commonly used as a screening module of the fetuses to detect in advance possible fetal problems that could result in irreversible neurologic damage or even fetal death during labour. Although it has been proven to be a useful tool for the obstetricians, suspicious fetal heart rate patterns lack specificity and false positive fetal heart rate traces may result in unnecessary intervention increasing the caesarean section delivery rate. More recently, non-invasive techniques such as reflectance pulse oximetry with continuous recording of functional oxygen saturation of fetal arterial blood during active labour appear to develop into an important additional source of information about fetal status, especially in cases of non-reassuring fetal heart rate patterns.
In recent years, several attempts have been made to automate the diagnosis of the fetal status. Computerised algorithms, artificial neural networks and hybrid architectures have been developed and validated in order to assess the fetal heart rate parameters (baseline of the fetal heart rate, accelerations and decelerations etc.). Our method is based on the heart rate variability analysis and on the statistical analysis of fetal pulse oximetry (FSpO2).
It has been evident that there is a significant relationship between the autonomic nervous system and cardiovascular function even before birth1. The development of quantitative markers of the autonomic activity has been encouraged by experimental evidence and heart rate variability represents one of the most promising such markers. Many commercial devices now provide automated measurement of heart rate variability, providing the physician with a seemingly simple tool for both research and clinical studies.
The clinical relevance of heart rate variability was first appreciated in 1965 when Hon and Lee2 noted that fetal distress was preceded by alterations in interbeat intervals before any appreciable change occurred in the heart rate itself. Later on, others focussed their attention on the existence of physiologic rhythms embedded in the beat-to-beat heart rate signal3–6. In 1981, Akselrod et al.7 introduced power spectral analysis of heart rate fluctuations using fast Fourier transform and pointed out the relation between the activities of the autonomic nervous system and the low and high frequency peaks of the frequency domains. Since then, frequency analysis of heart rate fluctuation has been performed widely8 and applications to fetal distress have been attempted9–12.
Most of these studies are based on the application of fast Fourier transform and autoregressive modelling. However, these algorithms have limitations in the study of long term non-linear variations of heart rate as well as in the analysis of transient alterations of heart rate. Since the fetal heart rate shows a long term non-stationary behaviour13, the application of the above methods is not very effective. Wavelet analysis14–20 has proved to be one of the most successful techniques for the analysis of signals at multiple scales, even when non-stationarities are present, which often obscure such signals21,22 and has rendered many successful applications in the area of biomedical signal processing23,24. Ivanov et al.25 used wavelet transform to study the temporal fluctuation of the high frequency component of the heart rate fluctuation. Later, Thurner et al.26 used a similar procedure and focussed on the values of the wavelet coefficients variance rather than on the scaling exponent of the wavelet transform.
There is evidently a need to improve intrapartum fetal surveillance. The ideal system of fetal monitoring, one that is safe, direct, continuous and non-invasive, with acceptable sensitivity and specificity has yet to be determined. Pulse oximetry has been used extensively in the fields of neonatology, adult intensive care and anaesthesia. Considering its successful application in these fields, it would seem logical to extend its use in the area of intrapartum fetal surveillance. Much research has gone into evaluating fetal pulse oximetry, its safety, accuracy and reliability in predicting neonatal outcome. There have been several studies conducted correlating the fetal heart rate patterns and the FSpO227.
In this study, we used the wavelet transform of the fetal heart rate variability and we applied self-organising map in order to investigate the relationship between the fetal heart rate variability in different scales and FSpO2 (taking as a threshold value for the FSpO2 measurement the 30% level and calculating the distance between the threshold and the minimum value of FSpO2 during a 10-minute segment) for normal and acidemic fetuses during the second stage of labour, which can be used to discriminate acidemic fetuses from normal ones. Our system accepts as inputs the fetal heart rate recording and the FSpO2 measurements. Preprocessing and artefact elimination takes place. Wavelet analysis is applied on 10-minute fetal heart rate segments and certain parameters (standard variation of the wavelet coefficients) are estimated, which together with the FSpO2 measurements comprise the inputs to the self-organising map neural network. The flow chart of the analysis is shown in Fig. 1.
Data were collected randomly from 61 women during labour, which took place in the Maternity Unit of the 2nd Department of Obstetrics and Gynaecology of the University of Athens. All women gave informed consent to the study, which was also approved by the Ethics Committee of the Aretaieion Hospital. The women were carrying singleton pregnancies of more than 37 weeks of gestation. The presentation was cephalic, and they were monitored after spontaneous or artificial rupture of membranes when cervical dilatation was more than 3 cm. Women with antepartum metabolic or endocrine disorders were not included in the study. Twelve cases, in which umbilical artery pH was lower than 7.20 and Apgar score was ≤7 at 5 minutes were grouped together in the risk group. The rest of the women formed the normal group. The cardiotocogram and the FSpO2 have been recorded during labour using the Corometrics Series 120 Cardiotocograph combined with fetal pulse oximeter. The duration of the second stage of the labour ranged from 20 to 50 minutes.
The fetal heart rate was measured externally. A transducer placed on the mother's abdomen is used to direct an ultrasonic beam toward the fetal heart and to sense Doppler shifted echoes created by moving cardiac structures. The sampling frequency was 1 Hz.
The percentage of the functional oxygen saturation of fetal arterial blood (FSpO2) was measured non-invasively by applying the Nellcor Puritan Benett fetal oxygen sensor to the cheek/temple area of the fetal head.
The second stage of labour of each case was divided into 10-minute segments (600 values of fetal heart rate). Artefacts and ectopic beats have been manually removed from the fetal heart rate signal and the FSpO2 recordings. Abrupt changes of FSpO2 were removed and linear interpolation was employed, when the duration of the artefact was below a certain value.
We applied wavelet multiresolution analysis for each 10-minute segment in order to address the problem of long term non-stationary behaviour of the fetal heart rate tracings and to estimate the power in different frequency ranges. The multiscale feature of the wavelet transform allows the decomposition of a signal in to a number of scales, each scale representing a particular ‘coarseness’ of the signal under study15. This essentially decomposes the signal into a set of signals of varying ‘coarseness’ ranging from low frequency components progressively to high frequency components. Thus, if one can make a decision concerning the underlying frequency components of the signal, one may choose the appropriate scale in the wavelet transform (smaller scales correspond to more rapid variations and therefore to higher frequencies), while ignoring the contribution of the other scales. This decomposition of the signal into different scales is particularly useful if the wavelet transform is performed on an orthogonal basis.
Wavelet analysis transforms the heart rate sequence into a space of wavelet coefficients. Technically, the coefficients are obtained by carrying out a discrete wavelet transform:
where the scale variable m and the translation variable n are non-negative integers, HR[i] is the discrete index sequence of the heart rate, ψ is called the mother wavelet function and M represents the total number of heart beats analysed.
Because certain wavelets (ψ) have vanishing moments, polynomial trends in the signal are automatically eliminated in the process of wavelet transformation. Since the signal fluctuates in time, so too does the sequence of wavelet coefficients at any given scale, although its mean is zero. A simple measure for this variability is the wavelet coefficient standard deviation, as a function of scale. This quantity has recently been shown to be quite valuable for heart rate variability26 and is related to the power spectrum density in a bandwidth surrounding the frequency fm corresponding to scale m28.
In the present study, the used mother wavelet was Daubechies 20-tap and the wavelet decomposition was performed up to scale index 4, as illustrated in Fig. 2. For each scale, we have calculated the wavelet coefficient standard deviation σwav(m), where m (1:4) is the scale index corresponding to the centre frequencies 0.34, 0.17, 0.08 and 0.04 Hz, respectively. σwav(1) was excluded from further analysis, since no significant difference was noticed between the normal and the risk groups.
For each 10-minute segment, we calculated the percentage of time in which the FSpO2 was less than 30% (SpO230).
We also calculated the distance between the lowest value of FSpO2 in the 10-minute segment and the 30% threshold (dist(30-minute FSpO2)).
In order to categorise the different 10-minute fetal heart rate patterns and the associated 10-minute FSpO2 segments, we used the self-organising map neural network with the Kohonen learning rule29. Such a network consists of two layers: an input layer and a two-dimensional output, Kohonen layer. Self-organising maps, also called topology-preserving maps, assume a topologic structure among the cluster units. They learn to recognise groups of similar input vectors, in such a way that neurones physically near each other in the neurone layer respond to similar input vectors. During the self-organisation process, the cluster unit whose weight vector matches the input pattern most closely (typically, the square of the minimum Euclidean distance) is chosen as the winner. The winning unit and its neighbouring units (in terms of the topology of the cluster units) update their weights. In this way, a mapping process takes place; input data vectors with similar features are mapped into the same area of the self-organising map.
The input vectors are the following: σwav(2), σwav(3), σwav(4), SpO and dist(30-minute FSpO2). After training the self-organising neural network, we applied the k-means clustering algorithm in order to estimate the number of the resulting clusters, and we calculated the projection matrices that indicate the values of the corresponding input parameters in each cluster at the two-dimensional output layer.
Figure 3 visualises the resulting clusters and the projection matrices of the input parameters σwav(2), σwav(3), σwav(4), SpO230 and dist(30-minute FSpO2). In cluster 1, σwav(2), σwav(3) and σwav(4) remain low and the SpO230 is minimal. In clusters 2 and 3, the measured standard deviations of wavelet coefficients are increased in relation to cluster 1 but the main differences between clusters 2 and 3 are:
(a)the σwav(4) is statistically significant lower (P < 0.05) in cluster 3 compared with cluster 2.
(b)the percentage of time in which FSpO2 is less than 30% is very high in cluster 3 compared with the other clusters.
(c)the distance between the lowest value of FSpO2 and the 30% threshold (dist(30-minute FSpO2)), in each 10-minute period is high in cluster 3 compared with the other clusters.
Table 1 shows the mean and standard deviation of the input parameters for the three different clusters.
Table 1. Input parameters for the different clusters. Values are given as mean (SD).
Cluster 3 attracts all the cases of the risk group except two and includes also one case of the normal group. The two cases of the risk group that were not included in cluster 3 were mapped in cluster 2, close to the borderline between clusters 2 and 3. The outcome of the cases that had at least one 10-minute period in cluster 3 classified them as risk group cases.
In order to characterise the performance of the neural network, we calculated the sensitivity and specificity for different pH values of the umbilical artery (see Table 2) and the subsequent receiver–operator characteristic curve shows that the cut off point of the system for the recognition of risk group cases was pH values <7.20 (Fig. 4).
Table 2. Sensitivity and specificity for different pH values. Values are given as %.
We have applied multiresolution wavelet analysis in order to address the problem of long term non-stationary behaviour of fetal heart rate. The calculation of the standard deviation of the wavelet coefficients enables us to extract information about the power spectrum density in a bandwidth surrounding the frequency fm corresponding to scale m, which in combination with the fetal pulse oximetry is used as input to the self-organising map neural networks for the categorisation of the fetal heart rate patterns.
Heart rate changes in a complex way due to several frequency domains. Until recently, analyses on the frequencies have been used to obtain information upon which a system is responsible in determining the heart rate. Regarding the frequency ranges that correspond to different physiologic alterations of the adult heart rate, it is shown that a very high frequency component between 0.75 and ∼1.5 Hz is the domain of heart beat, a high frequency component between 0.15 and ∼0.4 Hz is the domain of respiration due to the vagal nerve activity, a low frequency component between 0.04 and ∼0.15 Hz is the domain indicating the sympathetic and parasympathetic system interaction following baroreflex effect and the very low frequency component with the frequency under 0.04 Hz is known to play an important function in hormonal changes and thermoregulation although its precise origin is unknown30,31.
Power spectral analysis provides an index of circulatory rhythmicity. The frequency of the oscillations detected by power spectral analysis depends on inputs to the central nervous system with consequent alterations in efferent cardiac vagal and sympathetic activity. By blocking these efferent pathways, one should be able to determine the extent to which the blocked pathway normally influences the fetal heart. Based on this method, which was carried out in chronically catheterised fetal sheep32, it has been shown that under normoxaemic resting conditions, power spectral density in a frequency range between 0.04 and 1.3 Hz was strongly reduced by cardiac vagal blockade, but unaffected by β-adrenoceptor blockade. Thus, it would seem that power spectral analysis based on fast Fourier transform is a method of poor sensitivity for assessing cardiac sympathetic tone. This could be because fetal cardiac sympathetic discharge is less rhythmic31.
Moreover, they noticed that the increase in power spectral density at 0.04–0.45 Hz induced by hypoxaemia was abolished by vagal blockade with atropine. Thus, the cardiac vagal effect was responsible for these changes in the fetal heart rate variability spectrum during hypoxaemia. In addition, baro- and chemoreflex-related oscillations in heart rate have a frequency of 0.07 Hz in neonates33.
In our study, we have noticed that variability at scale 4 (0.04 Hz) in cluster 3 showed statistical significant reduction compared with cluster 2, in combination with increased percentage of time in which FSpO2 was less than 30%.
Based on the previous data regarding the frequency ranges that represent better the sympathetic and parasympathetic system activity, we could assume that in cluster 3 there is a predominance of vagal system activity while sympathetic action is reduced.
Fetal pulse oximetry seems to be an important additional source of information. Not only the time in which the FSpO2 was less than 30%, but also the distance of the minimum value of FSpO2 and the 30% threshold play an important role in the classification of the patterns, especially in the cases of non-reassuring fetal heart rate patterns.
We believe that computerised analysis of the fetal heart rate monitoring and pulse oximetry recordings based on the combination of wavelet analysis and artificial neural networks is a very promising technique in objective intrapartum diagnosis of fetal hypoxia. Further evaluation of this technique is mandatory to evaluate its efficacy and reliability in interpreting fetal heart rate recordings. Thus, we are going to increase the cases included in the study. We are currently analysing the data from the first stage of labour and further report is going to be available soon.
Our aim is to develop an on-line system that will recognise early the fetal hypoxia and alert the clinician to decide under objective conditions when and how to perform the delivery.
This study has been supported by the program ‘PENED’, General Secretariat of Research and Technology.