New Methods for Ctg Fetal Monitoring

Antepartum fetal monitoring based on the classical cardiotocography (CTG) is a noninvasive and low-price tool for checking fetal status. Its introduction in the clinical routine limited the occurrence of fetal problems leading to a reduction of the precocious child mortality. Nevertheless very poor indications on fetal pathologies can be inferred from the actual CTG analysis methods, either they consist of the clinician eye inspection or of automatic algorithms. It is certain that fetal heart rate and uterine contraction carry much more information on fetal state than it is extracted by classical analysis methods. In particular Fetal Heart Rate (FHR) signal has demonstrated to provide consistent indication of his well being status and in case of fetal stress, during labor, the FHR usually shows some morphological alterations. As the methods actually used for judging a CTG trace as "abnormal" give a too low predictive value for fetal dangers, we started to develop a new computerized system for the CTG analysis. The fetal monitoring system is based on a new multiparametric analysis of FHR which includes non-linear analysis algorithms (Approximate Entropy and space state maps) of FHR. The analysis is coupled with a classification of fetal states (ABCD) by means of Neural Networks. A comparison between supervised and unsupervised networks has been done on the same set of recordings. A prototype of this new monitoring system will be implemented on the basis of HP Traceview distributed architecture. INTRODUCTION The introduction of cardiotocography (CTG) as a non-invasive technique for monitoring fetal conditions, allowed obstetricians to direct their attention to the antepartum period, on the basis that a major portion of the unfavorable fetal outcomes seems due to events that occur prior to the onset of labour (van Gejin, 1996). As a matter of fact, a number of risky conditions for fetal compromise has been recognized in the antepartum period, of which intrauterine growth retardation (IUGR) due to uteroplacental insufficiency and maternal type I diabetes are the predominant. Since its introduction in the clinical routine, the use of CTG for antepartum fetal monitoring has led to a drastic reduction of intrapartum and precocious child mortality. However the actual methodology employed for judging CTG tracings has demonstrated a low predictive value for the fetal danger and high value of false positives in most cases. Indeed in the last 25 years since the introduction of CTG analysis, although the fetal and neonatal death rates have fallen considerably, the risk and severity of neurological handicap may even be rising. The conclusion of a number of recent studies is that very poor indications about fetus/newborn illness could be inferred from the actual CTG analysis (van Gejin, 1996), (Dawes et al. 1996). The most reliable indicator of fetal condition is represented by fetal heart rate (FHR) signal, upon which CTG is based. In case of fetal stress there is a high probability (>90%) that FHR, during labour, will show some anomalies or alterations, while if FHR recording seems to be normal, chances are high that the fetus can stand the labour. Several conditions such as hypoxia, acidemia, drug induction produce noticeable variations of FHR, which are usually detected by simple eye inspection of the physician. In essence, the main characteristic of the FHR is thought to be the presence of a baseline (sinusal rhythm) on which the frequency control mechanisms act by provoking some irregularities called accelerations and decelerations. (Mantel 1990). Up to now CTG records have been analyzed by detecting and classifying mainly the changes of that hypothetical FHR base value (accelerations and decelerations) in the hope of revealing a fetal sufferance status. Some CTG systems (Sonicaid, Hp) have tried an automatic classification of fetal states, based on the attempt to reproduce the criteria used by the clinicians, although in a quantitative way. In addition to the identification of accelerations and decelerations, also a quantitative assessment of the short term (STV) and long term (LTV) variability has been performed by the computerized CTG diagnostic systems. However the algorithmic approach, implemented on computerized automatic CTG diagnostic systems, has only led to a reduction of inter and intraobserver variability. An EEC project (Perinatal Monitoring) was pointed out to test the only commercially available system (Sonicaid System 8000) with an initial multitrial research (van Geijn, 1993) and the results did not show a significant clinical improvement from the classic analysis by eye inspection to the automatic one. On the other hand, recent studies on HR variability signal of adult and newborn subjects emphasize that both linear and nonlinear effects contribute to the signal generation pattern (Signorini et al., 1992 and 1994). In addition it was noted that fetal distress was preceded by alterations in interbeat intervals before any appreciable change occurred in heart rate itself. If observed on long periods of time, the series obtained from the HR values are highly irregular, typical of nonlinear system behavior. Moreover, using even nonlinear analysis techniques can cluster pathological conditions. All these results lead to think that FHR regulation mechanisms show an intrinsic nonlinear behavior, i.e. FHR values can highly oscillate in time and not always tend to an equilibrium state or to a sinusal rhythm. Thus, FHR variation contains the information about the neural events controlling fetal heart, although the methodological tools used for clinical diagnosis up to now did not allow to extract reliable quantitative indexes linking physiopathological fetal states with FHR signal patterns. For that reason we decided to face the problem of extracting from FHR signal both a classification of FHR patterns through ANNs, which are known to behave as non-linear classifiers, and new indexes of the non-linear behavior of FHR, presenting high sensitivity with respect to normal and pathological fetal states. DEVELOPMENT OF A NEW FETAL MONITORING SYSTEM The project we are developing has the goal to realize a new system for monitoring fetal condition, based on an appropriate and reliable analysis of FHRV. The final product of this work will be a clinical instrumentation prototype whose main characteristics are illustrated in the following. The whole system will be implemented on a PC workstation on a Windows platform. Its medium/low cost is crucial for being used in most obstetrical units. We identify two main features that will be the kernel of a new instrument for the fetal cardiotocographic analysis. Most of these are yet implemented and have been tested on an annotated set of CTG signals as will be illustrated in the Results section. 1. System learning ability This feature is obtained by means of Neural Network (NN) techniques for the diagnostic classification of FHR patterns based on the new set of relevant parameters identified by the new mathematical tools illustrated in the following. NNs have the peculiarity of being able to learn how to perform nonlinear classification of the input vectors once trained with an appropriate set of examples. Depending on the network architecture the training can be supervised – i.e. the training set consist of couples of input and desired output vectors or unsupervised – the learning process needs only the input vectors. 2. Linear and nonlinear multiparametric analysis of FHR We evaluate the FHR characteristics by calculating linear parameters (Power spectral density estimation, variance), by extracting regularity parameters (Approximate Entropy) and by representing signal variations through Delay Maps. The new System Build-up has been carried out trough a modification of both the hardware and the software of an existing computerized CTG system (HP-2CTG) based on the HP-M1351A ultrasound fetal monitor. We started from the HP-2CTG system and we modified the FHRV sampling frequency, passing from the collection of 1 FHR value every 2.5 sec to 1 FHR value every 0.5 or 0.25 sec. Following the modification of the FHR sampling frequency, the software was adapted for extracting, in a reliable way, the standard parameters (such as FHR accelerations, deceleration, etc.). Then, a first step for providing the system with learning ability was the implementation of two different NN architectures for the classification of FHR patterns based on input vectors consisting of 15 parameters automatically extracted by the 2CTG system. Once the NN has reached a stable state, the obtained clusters will be analyzed on the basis of the actual pathophysiological knowledge of FHRV indexes. The NN will then be tuned with the cases collected in the clinical tests in order to identify precise pathological classes. The final setup consists of a supervised NN, which will be implemented on the prototype system and trained with the clinical cases collected during the project. This NN can be re-trained by the end-user with his own cases. The second step consisted of the extraction of linear and nonlinear FHR global indexes including variance and Approximate Entropy. ApEn parameter quantifies the amount of regularity in data performing a detection of differences in HR that are not singled out by other classical analysis. Higher ApEn values indicate greater randomness in HR pulse. In this way, the normal fetal development should be characterized by an increasing of irregularity in HRV. As a matter of fact the complexity and the regularity properties of FHR dynamics can be useful to classify pathological situation as obtained for adult subjects and newborn infants. On the fetal data obtained from this system we compute of the power spectral density (PSD) of the FHR by means computation based on the autoregressive modeling approach. PSD analysis provides tools for a better identification of heart rate patterns related to the Autonomic Nervous System control activity. The construction of delay maps can reveal different level of signal complexity. It could become a simple but immediate tool for physicians quantifying differences in fetal patterns. Periodic and therefore predictable patterns could be distinguished from non-periodic and unpredictable ones. FHR behavior will be evaluate by this set of quantitative parameters in time and frequency domain. They are sensitive to different pathological states as it has been demonstrated for several HRV signal conditions (Task Force ESC-NASPE, 1996).