Validarea unui sistem de inteligenţă artificială pentru interpretarea traseului antepartum al ritmului cardiac fetal (Ankara System Version 5.0)

Introduction

Non-stress tests (NST) are widely used for the assessment of fetal well-being, mostly after 28 gestational weeks, within the framework of perinatal surveillance programs^(1,2). However, the evaluation of this test is full of subjectivities and, therefore, physicians are searching for better approaches and computerized methodologies^(3-8).

Various types of computerized systems have been developed for the evaluation of antepartum fetal heart rate (FHR) tracings^(3-9). However, there are some discrepancies behind NST that greatly affect its suitability for testing fetal well-being. These discrepancies include: the name of the test, phrases used for defining test results, standardization of applications (duration, timing, frequency etc.), differences in the algorithms of different “fetal monitors/transducers”, distinctions in terms of the manual evaluation of tracings, and the presence of different types of computerized systems that use different algorithms at different steps of the analysis and decision-making procedures, such as data acquisition, waveform analysis, feature extraction, classification and decision making^(3-10). Furthermore, the presence of different types of decision-making models, such as knowledge-based systems, supervised or unsupervised artificial intelligent systems, and hybrid systems in various forms, are all making antepartum FHR monitoring more complex and debatable^(3-9).

In this work, we have decided to present the validation results of our hybrid artificial intelligent system (version 5.0 prototype) for the interpretation of antepartum FHR tracings.

Patients/materials and method

This study consisted of 3100 assuring FHR tracings obtained from normal pregnancies and 467 non-assuring tracings (13.09% of all cases) obtained from pregnancies with various types of maternal health risk factors and/or obstetric complications. Assuring tracings from complicated pregnancies and non-assuring tracings from normal pregnancies were not included in the study, in order to maintain a qualified set of data. All tracings were evaluated twice by a single expert. The first evaluation took place just after data acquisition and the second occurred two weeks later, without knowledge of the clinical characteristics of the patient (unmatched tracing evaluations were not included in the database). A total of 3032 tracings were used for the training of the algorithms, while 535 were used for testing and validation.

Fetal heart rate tracings were collected during the antepartum period, from the 28^th gestational week onwards. A Doppler ultrasound heart rate monitor with autocorrelation (Hewlett Packard 8040A) was used to detect fetal pulse. Perceived fetal movements were recorded by the patient. The fetal monitor was connected to the computer system by a special frame grabber. Ten-minute recordings were kept in a special file for pattern recognition and decision-making.

Wavelet transformation is used for waveform analysis to extract 100 features from each tracing which was the first step of pattern recognition (feature extraction)^(11,12). Then, artificial neural network (ANN) was used for the classification of waveforms as the second step of pattern recognition^(13-15).

Firstly, unsupervised ANN (clustering by vector quantization) (Step IIa) was used to define six clusters of tracings, using the extracted 100 features^(14-16). Then, the medical expert evaluated the member tracings of each cluster in terms of presence of assuring, non-assuring and unidentified waveforms^(1,10). Clusters that were dominant in terms of assuring and non-assuring tracings were selected for the next step (Step IIb). Subsequently, tracings of these clusters were reused for the second classification by a supervised ANN (back propagation learning algorithm); the tracings were classified as assuring or non-assuring FHR waveforms as a final decision⁽¹⁷⁾.

This study is supported by the State Planning Organization of Turkey (Project Numbers: DPT 90K 120540 and DPT 96K 120540) and The Scientific and Technological Research Council of Turkey (TAG-595).

Results

The testing group consisted of 480 assuring and 55 non-assuring tracings.  Table 1 shows the matrix related to the validation and statistical measures of this version of the system. The sensitivity and specificity of Ankara System (Version 5.0) were found to be 61.25% and 98.68%, respectively. The positive and negative predictive values of the system were 89.1% and 93.5%, respectively.  Table 2 shows the validation and the statistical measures of this version of hybrid ANN.
 

Discussion 

Antepartum FHR tracing is used in different forms in perinatal surveillance, such as the use of the NST^(1,2,9). However, the NST or however it is referred to (even as “antepartum FHR testing”) has some disadvantages in terms of standardization problems. The factors that make NST more complex include the variable duration of recordings, the presence of different fetal monitors with different algorithms, and autocorrelation-related differences. In classical manual and observational practice, various indices and parameters are used for the evaluation of the test; these include baseline, variability, FHR acceleration, FHR deceleration and perceived fetal movement^(18-20). However, we do not know the exact biological rationales behind the condition-related changes of these variables. It might be concluded that manual interpretation of the tracing (NST) is subjective and consists of various types of heuristic approaches. The sensitivity and specificity of NST are reported to be around 50% and 98%, respectively^(1,2,9).

Various computerized systems have been developed to overcome the problems inherent in manual evaluations of FHR tracings (NST)^(3-8). Then, however, we face the standardization problems of these computerized systems. In our institution, we have started with rule-based automated systems and used various types of algorithms for baseline, variability and other parameters^(3-5). However, a lack of sufficient information behind the condition-related changes of these parameters has directed us to develop an artificial intelligent system.

Artificial intelligence systems are widely used in different fields of medicine for a variety of purposes^(21-27). In this study, we have used pattern recognition (feature extraction and classification) and decision-making algorithms to develop our hybrid artificial intelligent system (Ankara System Version 5.0). Wavelet transformation is used for feature extraction^(12,13). Then, the extracted 100 features were used for the classification steps^(16,17). In the first step of classification, an unsupervised ANN (vector quantization) was used for clustering. In the second step of the classification, we used a supervised ANN (backpropagation learning algorithm) and thus classified the tracings as either assuring or non-assuring. The sensitivity and specificity of the Ankara System (version 5.0) were found to be 61.25% and 98.68%, respectively. We believe the statistical measures of this version are sufficient enough and may be used, as well as classical NST, which is especially important in the perinatal surveillance of high-risk pregnancies.

Caron et al. developed a system for the computerized assessment of fetal behavioral states in order to guide the physicians objectively for the management of pregnancies, with promising results⁽⁸⁾. Similarly, Hiett et al. used the Oxford Sonicaid System 8000 for comparing visual versus computerized analysis of non-stress tests⁽⁶⁾. Non-stress tests of 575 high-risk patients were analyzed by two experts and the computerized system, without knowledge of the pregnancy outcomes. Similar rates of sensitivity, specificity, and positive and negative predictive values were found in both approaches. Thus, the authors concluded that automated fetal heart rate testing might become an acceptable alternative to conventional visual analysis⁽⁶⁾.

Furthermore, Joana Lopes-Pereira et al. conducted a retrospective cohort study including 38,466 deliveries to evaluate the performance of the Omniview-SisPorto system in assisting physicians for the interpretation of intrapartum cardiotocography traces⁽⁷⁾. A significant reduction in the incidence of hypoxic-ischemic encephalopathy and a modest reduction in caesarean deliveries were observed⁽⁷⁾. However, there are some studies in the literature with conflicting results^(9,28). Two recent examples of meta-analyses did not report favorable outcomes with the application of computerized systems. However, these studies had some limitations due to inclusion criteria, risk of bias and study design^(9,28).

We believe that these types of decision-making technologies and models are important in order to produce better results and increased objectivity in the evaluation of FHR tracings^(3-9). Another advantage of “expert and the artificial intelligent systems” is the possibility of producing institute-based versions (in terms of expert physicians and hardware). However, there are still standardization problems related to methodological and technological differences. In this version, we chose to use single expert FHR evaluations at the training phase, in order to maintain uniformity for the final interpretations. We believe that our approach will stimulate further studies in this field.

In conclusion, the methodologies used in the creation of a hybrid artificial intelligent system (Ankara System Version 5.0) might be an alternative for future computerized systems to be used in the interpretation of FHR tracings.  

Acknowledgments: Special thanks to all perinatology staff who work with devotion for improving the health standards of the patients.

Funding: No funding was used for this study.

Conflict of interests: The authors declare no conflict of interests.