5.2 Clinical Data
5.2.4 Data annotation: fetal QRS annotation
Another feasible analysis using the available clinical dataset is to assess the accuracy of the attained FQRS detections. Exact FQRS are crucial not only for enabling FHR/FHRV analysis but also because morphological analysis of the FECG signal usually relies on FQRS locations to segment the FECG waves. Using a clinical database, one is able to validate the results obtained in a simulated dataset (e.g. FECGSYNDB). Therefore, a conclusive assessment of how these preprocessing, extraction, and detection methods perform is possible.
Table 5.8:Correlation table for clinical data regardingW OG <28. Asterisk notation show the significance level as significant (∗), very significant (∗∗) and highly significant (∗∗∗). Abbrevia- tions used: position (pos.), consensus (cons.) and visibility (vis.).
(a) Effect size for interval-ordinal and ordinal-ordinal variables (both usingρ)
Parameter WOG† amniotic
fluid cons.SNR FECGcons. amniotic fluid −0.28∗∗∗ – – –
SNR cons. ׇ −0.02 – –
FECG cons. ׇ −0.03∗ 0.14∗∗∗ –
vis. cons. ׇ −0.03∗∗ 0.18∗∗∗ 0.99
† Before calculating the effect size, repeated terms were excluded.
‡ Don’t-care term. Since every channel number is contained in every recording,
these correlations shall be further investigated using the RI.
(b) Effect size for nominal-nominal (w) and nominal-ordinal/interval (η) variables
Parameter channel # WOG† placenta
pos. amnioticfluid fetal pos.
WOG ׇ – – – – placenta pos. ׇ 0.99 – – – amniotic fluid ׇ – 0.07∗∗∗ – – fetal pos. ׇ 0.99 0.42∗∗∗ 0.25∗∗∗ – SNR cons. 0.34 – 0.76∗∗ – 0.79∗∗∗ FECG cons. 0.78∗∗∗ – 0.30∗∗∗ – 0.23∗∗∗ vis. cons. 0.76∗∗ – 0.21∗∗∗ – 0.14∗∗∗
Hence, a subset14of 24 recordings of viable quality were annotated for MQRS and FQRS locations. These recordings were taken from ten women (both healthy and pathological patients were present), aging from 21 to 33 years (27.1±4.3 years), gestational weeks between 20 and 28 weeks (25.0±2.5 weeks) and duration of 19.4±2.4 minutes, where no significant arrhythmia or ectopic beats have been found for neither mother or fetus. Considering the signal quality annotations presented in the previous section, the selected recordings had different RI indices: six recordings withRI = 2, eleven withRI = 3, two withRI = 4 and five withRI = 5. Each recording had its MQRS and FQRS annotated by one and corrected by two other trained specialists. Annotators should use four types of annotations for “visible peak”, “likely, yet, not visible peak”, “begin” and “end of signal loss interval”. The annotation procedure was performed using a JAVA GUI developed at the IBMT. Both software and annotation protocol are detailed in AppendixB. The annotated data totalized 465 min, containing over 67,000 fetal complexes, and was used as gold-standard for evaluating the developed FQRS/FHR detection techniques.
14The reason behind the choice for a subset was due to the fact that the annotation procedure is very time-costly.
The annotation of a single recording by one expert can take a couple hours. The author assumes if the subset is heterogeneous enough, it may nonetheless provide a good reference for further benchmarks.
5.3 Chapter Summary
In this chapter, the two principal sources of data developed throughout this work were presented. First, motivated by the lack of freely available abdECG databases, in cooperation with the IBME (University of Oxford) a NIFECG simulator was developed, i.e. the so-called FECGSYN (see Section5.1.1). This versatile toolbox provides means for several types of non- stationary abdominal mixtures and several pathophysiological scenarios. By making use of the FECGSYN, a large open-access database was created (i.e. FECGSYNDB – see Section5.1.2). Second, together with the project partners at the University Hospital of Leipzig, a clinical trial was conducted and resulted inn= 259 NIFECG recordings. The obtained clinical data was statistically characterized in Section5.2. The overall quality of NIFECG at early WOGs (i.e. <28 weeks) was shown to be fair. Particularly in this analysis, the dependence between WOG and FECG amplitude was demonstrated. Indeed there is a strong attenuation of the FECG around the 28-32 weeks, which is commonly associated with the vernix caseosa. Another important factor discussed was the influence on electrode location and the power of fetal and noise signals, where depending on the NIFECG application, a mixture of more and less distant electrodes may be considered. Other influencing factors were also evaluated using a bivariate analysis, however, the effects are not conclusive due to the modest size of the clinical dataset. Further analysis should focus on multivariate analysis using a much larger dataset (i.e. at least a couple thousand recordings). In the next chapter, the methods proposed in Chapter4 are comprehensively benchmarked against several other extraction methods available in the literature, using the produced data material here presented.
6
Results for Data Analysis
In the previous chapters, several improvements on current extraction and detection tech- niques for NIFECG analysis as well as valuable data were produced. In this chapter, these suggested approaches are benchmarked in depth. In order to do so, the present chapter is divided into two main sections, based on the type of analysis performed and data material used. In Section6.1, the simulated data (i.e. the FECGSYNDB - see Section5.1.1) was used to benchmark several extraction methods, in terms of FQRS detection and morphological fidelity of NIFECG estimates. In Section6.2, the best performing methods from each algorithmic category (see Figure3.4) were applied to the own clinical dataset collected throughout this work (see Section5.2).6.1 Simulated Data
The application of simulated data for the purpose of characterizing NIFECG is particularly important, since one can evaluate the different algorithms behavior in the presence of differ- ent noise levels, non-stationary artefacts and pathophysiological events. In this section, the FECGSYNDB was used to that end in an analog manner as in Andreottiet al.[18]. This database comprises several pathophysiological scenarios, shown in Table5.1.