Non-invasive fetal ECG (fECG) is a promising technology for long-term monitoring of pregnancies and thus the early detection of fetal abnormalities. One issue that occurs due to the varying signal quality are detection errors. In this work, we established a novel measure of the detection quality for fetal ECG beat detections and refer to them as detection quality indices (DQI). The aim of this study was to develop a methodology which identifies detection errors and can deal with them.
We categorized the DQIs in four independent categories of information: ration of fetal-maternal heart rate, the similarity of fetal QRS and maternal QRS detections, the extravagance of detection, and morphological information and reviled different behaviours of the classes concerning the different categories.
With the DQIs considered as features we trained a decision tree as an interpretable model that can sharply distinguish between correct detected fQRS complexes and two detection errors (f1 = 0.99). This revealed a deeper understanding of the used detector and indicates a promising way to further improve the algorithm. The method was able to improve the error of time-domain heart rate variability parameters and approximated entropy of two measurements of one subject compared to manual detection by 35 % on average.
Our results sound promising because we just excluded up to 2 % of invalid beat detentions and not take further steps to find QRS complexes in the areas of exclusion. Therefore, we conclude that DQIs are useful to understand the behaviour of a QRS detector better and can improve the quality of QRS detections. Further studies should consider other detectors and examine the effect on HRV parameters in larger datasets.