Comparison and verification of detection accuracy for late deceleration with and without uterine contractions signals using convolutional neural networks

Cardiotocography (CTG) is the most common technique used to noninvasively record fetal heart rate (FHR) and uterine contraction (UC) over time to monitor and assess fetal health (Pinas and Chandraharan, 2016; Santo et al., 2017). CTG is used to determine the presence or absence of fetal heart deceleration and its type based on the temporal characteristics of the FHR and UC (Lear et al., 2016). One of the critical clinical prognostic signs of fetal outcome is late deceleration (LD), which is caused by a decrease in uteroplacental blood flow associated with uterine contraction, resulting in reduced gas exchange in the placenta. LD has been shown to represent a chemoreceptor-mediated response to fetal hypoxemia. In other words, LD occurs when fetal oxygenation becomes inadequate and the chemoreceptor threshold is exceeded. Consequently, the fetal heart rate gradually decreases and then gradually recovers (Itskovitz et al., 1982; Pinas and Chandraharan, 2016). Occasional LD occurring at less than 50% of uterine contractions has been previously described as an “early marker of fetal mild hypoxemia” (Sameshima et al., 2004), obstetricians should be informed as soon as possible about the occurrence of LD. Early detection of LD is critical because it can prompt close monitoring and preparation for emergency delivery. LD is evaluated visually by obstetricians according to guidelines such as the International Federation of Gynecology and Obstetrics (FIGO) and the Japanese Society of Obstetrics and Gynecology. While the FIGO defines LD as an amplitude of ≥15 bpm, the guidelines for obstetric practice in Japan define mild LD as an amplitude of <15 bpm and severe LD as an amplitude of ≥15 bpm; there are slight differences in the definitions between the guidelines.

The temporal relationship between FHR and UC is an important factor in determining LD; however, the signal quality of UC tends to be lower than that of FHR. External tocodynamometry measures uterine contraction using pressure changes obtained from a sensor attached to the maternal abdominal wall, and studies have indicated that the quality of UC is degraded when there is a large amount of abdominal fat or when the sensor is displaced by the mother’s body movements (Euliano et al., 2013; Vlemminx et al., 2017; Petrozziello et al., 2018). When the quality of UC signals is degraded owing to these causes, a situation arises in which the temporal relationship between FHR and UC, an important factor in determining LD, cannot be determined.

It is crucial that healthcare professionals identify the onset of LD as early as possible. Therefore, in clinical settings, when the quality of the UC signal is poor, attention is paid to the FHR signal pattern, and LD is evaluated based on the FHR signal alone. The morphological characteristics of LD include deceleration with a gradual onset, gradual return to baseline, and/or reduced variability within the deceleration (Ayres-de-Campos et al., 2015). Therefore, obstetricians may extract possible LD signals by capturing the morphological features of the FHR.

To detect the occurrence of LD more quickly without missing it, several efforts are underway to detect LD by automatically analyzing CTGs. Products using signal analysis to assist obstetricians in visual assessment include Omniview-SisPorto^Ⓡ 3.5 (Speculum, Lisbon, Portugal) and Trium CTG Online^Ⓡ (GE HealthcareVR, Little Chalfont, UK) (Magawa et al., 2021; Bernardes, 2023). Omniview-SisPorto^Ⓡ 3.5 is built on signal analysis algorithms compliant with FIGO guidelines, while Trium CTG Online^®, which is available in Japan, is designed based on the guidelines from the Japanese Society of Obstetrics and Gynecology. In a study comparing the analysis results of each product with visual evaluation by obstetricians, Costa MA et al. evaluated Omniview-SisPorto^Ⓡ 3.5 and reported a 68% agreement rate for decelerations (Costa et al., 2010). Contrastingly, Magawa S et al. evaluated Trium CTG Online^Ⓡ and reported a sensitivity of 0.93 and specificity of 0.99 for Severe LD (Magawa et al., 2021). A recent study examined the use of machine learning (ML) and deep learning (DL) to detect LD and compared the ML and DL methods (Das et al., 2023). This study reported an improved accuracy when labels determined using fuzzy logic were used as the gold standard. However, when the labels determined by the obstetricians were used as inputs for DL, the receiver operating characteristic (ROC) for LD detection was <0.60.

Although studies using signal analysis, ML, and DL have achieved notable results because both FHR and UC signals were used for analysis, it is thought that if the quality of the UC signal deteriorates, it would not be possible to determine the LD accurately. If LD can be detected when the UC signal is poor, it would be of great clinical significance because it would allow the early detection of LD without missing its onset. We also believe that more advanced signal analysis methods are needed to distinguish between LD and other conditions assessed using only FHR signals. Therefore, we constructed a method for detecting LD using convolutional neural networks (CNNs), which is one of the DL methods that has attracted attention in recent years in the field of obstetrics for research on evaluating the health of fetuses from CTG signals (Arain et al., 2023; Cömert et al., 2019; Hirono et al., 2024a; Ogasawara et al., 2021). Therefore, the objective of this study was to construct a robust model for detecting LD, even when the quality of the UC signal is poor, using a CNN to detect LD from the FHR signal only. The developed Artificial Intelligence (AI) model will enable faster and more accurate detection of LD and improve fetal prognosis.

This study was approved by The Institutional Review Board of the Niigata University of Health and Welfare (Approval No. 19397-241014). This section describes the experimental environment, data details, artificial intelligence (AI) models, and evaluation methods.

Figure 2 shows a flowchart for the ground-truth labeling and data selection. Two cases were excluded because the time from the start of recording to the onset of LD was too short to allow for data length, and one case was excluded because it was difficult to determine the LD within the data length determined in this study. As a result of the labeling and data selection, the final data used for analysis included 86 cases of LD, 440 cases of non-LD, and 23 cases of LD-like. To evaluate the generalization performance of the model, a 10-fold cross-validation was performed. For each dataset, the cross-validation, training, and validation datasets had a 9:1 ratio. Table 2 shows the numbers of LD and non-LD cases per fold. Due to a lack of the number of LD data, the data were augmented five times by shifting the data 30 s before and after the nadir point of the FHR and 15 s each. This made the number of LD and non-LD cases almost equal. By adjusting the detection accuracy to the highest value for each model, the FHR-only model showed the highest accuracy with four convolutional layers, and the FHR + UC model showed the highest accuracy with six convolutional layers. The kernel size was 1 × 13 for the FHR-only model and 1 × 17 for the FHR + UC model. The learning rate and the batch size for both models were 0.01 and 32, respectively.

Figure 3 summarizes the LD and non-LD classification accuracies in the FHR-only and FHR + UC models. Table 3 shows the F-measure for the LD and AUC for each model. The AUC and F-measure for LD were 0.896 and 0.613, respectively, for the FHR-only model and 0.928 and 0.711, respectively, for the FHR + UC model. Figure 4 shows the average output values of each model, with no statistically significant differences in the paired-samples t-test and overlapping error bars in the 95% confidence intervals of both models. Therefore, although the FHR-only model was less accurate than the FHR + UC model, there was no significant difference in the accuracy of LD detection, indicating that the FHR-only model also showed good results.

In the present ground-truth labeling, two expert obstetricians judged 4.2% (23/552 cases) of cases as LD-like. An example of a CTG waveform determined as an LD-like signal is shown in Figure 5. Table 4 and Figure 6 show the results of the validation using LD-like and non-LD cases. The FHR-only model yielded AUC and F-measures of 0.833 and 0.726, respectively, both of which were more accurate than the FHR + UC model. The 95% confidence interval error bars of the two models did not overlap in terms of the likelihood of LD, and the paired-sample t-test results showed that the FHR-only model was statistically superior in detecting LD-like cases (p < 0.001). Table 5 and Figure 7 present the validation results when LD-like was assumed to be LD and added to the LD cases. The FHR-only model yielded AUC and F-measure values of 0.867 and 0.747, respectively, both of which were more accurate than those of the FHR + UC model. The 95% confidence interval error bars of the two models did not overlap with the likelihood of LD, and the paired-sample t-test results showed that the FHR-only model was statistically superior in detecting LD-like cases (p < 0.001). Figure 8 shows case (A), which was most reliably detected as LD among the LD-like cases in the FHR-only model, which was highly accurate in detecting LD and LD-like cases, and case (B), in which a non-LD case was incorrectly detected as LD. Case (A) was highly accurate in detecting LD, and the morphological features of the FHR indicated that it could detect potential LD.

In this study, LDs were detected with high accuracy even in a model that used only the FHR signal. Therefore, the CNN constructed in this study could detect LD by capturing the baseline level of the FHR signal and its variability in acceleration or deceleration (Czabanski et al., 2012). The detection of LD using only the FHR signal may aid early, unmissable capture of signs of nonreassuring fetal status, even when the UC signal quality is poor. Additionally, inter- and intra-observer decipherment errors have been noted in the visual evaluation of CTG (Hernandez Engelhart et al., 2023; Rei et al., 2016); if the quality of the UC signal is poor, the difficulty in deciphering UC may lead to greater decipherment errors. Therefore, detecting LD with high accuracy and providing alerts based on FHR signals alone better supports CTG decipherment. Furthermore, alerting obstetric healthcare providers to signs of nonreassuring fetal status may encourage them to reattach sensors that measure pressure changes during uterine contractions. Reattachment of the sensor is of great clinical significance because it enables subsequent CTG readings to evaluate fetal health status in consideration of the temporal relationship between FHR and UC.

The results of the inference using the LD-like case showed that the FHR-only model could detect potential LD signals with high accuracy. Since it is important not to miss an LD, an LD-like signal should be detected as a potential LD signal. Thus, assuming that LD-like cases are indeed signs of LD and UC could not be detected, alerting them using the FHR-only model may lead to earlier detection of LD, the early marker of fetal mild hypoxemia.

In this study, we tested whether LD, the early marker of fetal mild hypoxemia, could be detected from only the FHR signal. However, specificity needs to be considered for clinical use. Particularly, since it is important to discriminate cases of variable deceleration (VD) (Japan Council for Quality Health Care, 2018), which is considered difficult to discriminate from LD and other decelerations, we conducted an evaluation using cases with VD. Of the non-LD cases in the database, 4.8% (21/440) were determined to have VD by the consensus of expert obstetricians at the time of final determination. The F-measure and percentage of correct answers for VD cases in each model were 0.822% and 66.6% (14/21), respectively, for the FHR-only model and 0.680% and 47.6% (10/21), respectively, for the FHR + UC model. The FHR-only model achieved higher accuracy in terms of both F-value and correct response rate. This suggests that the LD and VD could be detected with high accuracy using only the FHR signal. The differences in the morphological characteristics of the FHR in cases of the LD and VD are that the VD shows a rapid drop (onset to nadir in less than 30 s), good variability within the deceleration, and rapid recovery to baseline (Ayres-de-Campos et al., 2015). This suggests that the LD and VD could be correctly distinguished by capturing the morphological characteristics of the FHR. VD also accounts for the majority of bradycardia that occurs during labor and is considered a baroreceptor-mediated response to elevated arterial pressure, such as that occurring with cord compression (Ball and Parer, 1992). In other words, since VD is often a physiological response of the fetus during the delivery period, it is beneficial to distinguish it from LD, which is a sign of fetal hypoxemia.

Subsequently, 440 non-LD cases were analyzed. In 14.7% (65/440) of false positive cases, non-LD cases were misclassified as LD. Typically, 33 cases of intermittent interruption of the FHR signal were included. Additionally, three false positive cases were very close to the signal pattern of LDs. Specifically, in the case (B) shown in Figure 8, two expert obstetricians determined that the patient had PD because the recovery from the heart rate drop to baseline was only slightly longer than 3 min. Since PD is also a sign of fetal hypoxemia (Chandraharan and Arulkumaran, 2007), we believe that it is worthwhile to alert the observer in the same manner as that in LD; therefore, we do not see a major problem with false positives in these cases.

This study has three limitations. First, we used a simple CNN configuration model because the dataset was relatively small. After increasing the amount of data, we expect to obtain better results using the latest models. Second, we used fragmented data. It is necessary to evaluate this by constructing a system that can discriminate LD in real-time using continuous data. Third, we limited the target of detection to the first LD; however, we believe that it is important to detect LD as early as possible to ensure early medical intervention. However, it is possible to improve the detection accuracy of LDs if we use all the LDs in the dataset, not just the first.

This comparison and verification of the accuracy of LD detection with and without UC signals using a CNN showed that LD could be detected with high accuracy from the FHR signal only. This result suggests that the LD could be detected using the FHR signal, even when the quality of the UC signal is poor. Alerting obstetric healthcare providers to signs of nonreassuring fetal status will encourage them to closely monitor and prepare for rapid delivery, which is expected to improve fetal outcomes.

The author(s) declare that financial support was received for the research, authorship, and/or publication of this article. This work was funded by the JSPS Grant-in-Aid for Scientific Research (C) (Grant Number JP 24K13872). This work was supported by TOITU, Inc. IS received consulting fees from TOITU Co. Ltd. KY and SK have received a research grant from TOITU Co., Ltd.

The original contributions presented in the study are included in the article/supplementary material, further inquiries can be directed to the corresponding author.

IS: Conceptualization, Data curation, Formal Analysis, Funding acquisition, Investigation, Methodology, Project administration, Resources, Software, Supervision, Validation, Visualization, Writing–original draft, Writing–review and editing. YH: Conceptualization, Methodology, Software, Visualization, Writing–review and editing. ES: Conceptualization, Formal Analysis, Methodology, Visualization, Writing–review and editing. HY: Conceptualization, Formal Analysis, Methodology, Visualization, Writing–review and editing. KY: Conceptualization, Methodology, Visualization, Writing–review and editing, Funding acquisition. CK: Conceptualization, Methodology, Visualization, Writing–review and editing. AY: Conceptualization, Methodology, Visualization, Writing–review and editing. FU: Conceptualization, Visualization, Writing–review and editing. NK: Conceptualization, Visualization, Writing–review and editing. SK: Conceptualization, Data curation, Methodology, Supervision, Visualization, Writing–review and editing, Funding acquisition.

Authors YH and FU were employed by TOITU Co. Ltd.

The remaining authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.

The author(s) declare that no Generative AI was used in the creation of this manuscript.

All claims expressed in this article are solely those of the authors and do not necessarily represent those of their affiliated organizations, or those of the publisher, the editors and the reviewers. Any product that may be evaluated in this article, or claim that may be made by its manufacturer, is not guaranteed or endorsed by the publisher.

CTG, Cardiotocography; FHR, Fetal heart rate; UC, Uterine contraction; LD, Late deceleration; AI, Artificial Intelligence; CNN, Convolutional Neural Network; AUC, Area under the curve; FIGO, International Federation of Gynecology and Obstetrics; ML, Machine learning; DL, Deep learning; ReLU, Rectified Linear Unit; ROC, Receiver Operating Characteristic; TN, True negative; FN, False negative; TP, True positive; FP, False positive; VD, variable deceleration; PD, prolonged deceleration.