@techreport{oai:ipsj.ixsq.nii.ac.jp:00217997,
 author = {Xintong, Shi and Kohei, Yamamoto and Tomoaki, Ohtsuki and Yutaka, Matsui and Kazunari, Owada and Xintong, Shi and Kohei, Yamamoto and Tomoaki, Ohtsuki and Yutaka, Matsui and Kazunari, Owada},
 issue = {10},
 month = {May},
 note = {For fetal heart rate (FHR) monitoring, the non-invasive fetal electrocardiogram (FECG) obtained from abdomen surface electrodes has been widely employed. The accuracy of FECG-based FHR estimation, on the other hand, is highly dependent on the quality of the FECG signals, which can be influenced by a variety of interference sources, including maternal cardiac activities and fetal movements. Thus, FECG signal quality assessment (SQA) is a critical task in improving FHR estimation accuracy by eliminating or interpolating FHRs estimated from low-quality FECG signals. Various SQA approaches based on supervised learning have been proposed in recent studies. These approaches are capable of accurate SQA, however, they require big datasets with annotations. Nevertheless, the annotated datasets for the FECG SQA are extremely limited. In this research, to deal with this problem, we introduce an unsupervised learning-based SQA approach for identifying high and low-quality FECG signal segments. Some features associated with signal quality are extracted, including four entropy-based features, three statistical features, and four ECG signal quality indices (SQIs). In addition to these features, we introduce an autoencoder (AE)-based feature, which is based on reconstruction error of AE that reconstructs spectrograms obtained from FECG signal segments. The collected features are then input into the self-organizing map (SOM) to classify the high and low-quality FECG segments. The experimental results indicated that our approach classified high and low-quality signals with a 98% accuracy., For fetal heart rate (FHR) monitoring, the non-invasive fetal electrocardiogram (FECG) obtained from abdomen surface electrodes has been widely employed. The accuracy of FECG-based FHR estimation, on the other hand, is highly dependent on the quality of the FECG signals, which can be influenced by a variety of interference sources, including maternal cardiac activities and fetal movements. Thus, FECG signal quality assessment (SQA) is a critical task in improving FHR estimation accuracy by eliminating or interpolating FHRs estimated from low-quality FECG signals. Various SQA approaches based on supervised learning have been proposed in recent studies. These approaches are capable of accurate SQA, however, they require big datasets with annotations. Nevertheless, the annotated datasets for the FECG SQA are extremely limited. In this research, to deal with this problem, we introduce an unsupervised learning-based SQA approach for identifying high and low-quality FECG signal segments. Some features associated with signal quality are extracted, including four entropy-based features, three statistical features, and four ECG signal quality indices (SQIs). In addition to these features, we introduce an autoencoder (AE)-based feature, which is based on reconstruction error of AE that reconstructs spectrograms obtained from FECG signal segments. The collected features are then input into the self-organizing map (SOM) to classify the high and low-quality FECG segments. The experimental results indicated that our approach classified high and low-quality signals with a 98% accuracy.},
 title = {Unsupervised Learning-based Non-invasive Fetal ECG Signal Quality Assessment},
 year = {2022}
}