Multimodal Causal Machine Learning for Fetal Asphyxia Risk Prediction from Imperfect Monitoring Signals

Multimodal Causal Machine Learning for Fetal Asphyxia Risk Prediction from Imperfect Monitoring Signals | Research Square window.SnipcartSettings = { analytics: { enabled: false } }; (function() { var accessVector = localStorage.getItem('access_vector') || ''; window.dataLayer = window.dataLayer || []; if (accessVector) { window.dataLayer.push({ user: { profile: { profileInfo: { snid: accessVector } } } }); } })(); (function(w,d,s,l,i){w[l]=w[l]||[];w[l].push({'gtm.start':new Date().getTime(),event:'gtm.js'});var f=d.getElementsByTagName(s)[0],j=d.createElement(s),dl=l!='dataLayer'?'&l='+l:'';j.async=true;j.src='https://www.googletagmanager.com/gtm.js?id='+i+dl;f.parentNode.insertBefore(j,f);})(window,document,'script','dataLayer','GTM-K279D39R'); Browse Preprints In Review Journals COVID-19 Preprints AJE Video Bytes Research Tools Research Promotion AJE Professional Editing AJE Rubriq About Preprint Platform In Review Editorial Policies Our Team Advisory Board Help Center Sign In Submit a Preprint Cite Share Download PDF Article Multimodal Causal Machine Learning for Fetal Asphyxia Risk Prediction from Imperfect Monitoring Signals Debjyoti Karmakar, Lochana Mendis, Emerson Keenan, Marimuthu Palaniswami, and 2 more This is a preprint; it has not been peer reviewed by a journal. https://doi.org/ 10.21203/rs.3.rs-7122719/v1 This work is licensed under a CC BY 4.0 License Status: Posted Version 1 posted You are reading this latest preprint version Abstract Cardiotocography (CTG) is the cornerstone of intrapartum surveillance, yet its predictive value for perinatal asphyxia remains limited. Modern monitors record fetal heart rate (FHR) via Doppler ultrasound or intrapartum fetal electrocardiography (ECG), the latter usable only after cervical dilatation. We analysed 52,757 labours ≥36 weeks’ gestation across three hospitals—two in Australia and one in Europe—treating mode-specific FHR dropout (zero-value samples signifying missing valid biosignal) as an informative exposure rather than benign artifact. High dropout (>30% of the CTG recording available for analysis) was evaluated using doubly robust causal machine learning and survival analysis, incorporating inverse-probability weighting, targeted maximum-likelihood estimation (TMLE), and weighted Cox models. In the training cohort (n = 36,792), high ECG-mode dropout tripled the odds of perinatal asphyxia (OR 3.14, 95% CI: 2.20–4.48) and quadrupled the hazard (HR 4.76, 95% CI: 2.98–7.62) compared to ultrasound-mode low dropout. In crude analyses, this corresponded to an absolute risk increase of 5.0% (1 in 20 births). TMLE-adjusted models confirmed a 2.27% absolute risk increase—equivalent to one additional case per 44 labor episodes (95% CI: 28–83). External validation confirmed consistent associations in European (OR 4.02) and Australian (OR 2.22) cohorts. “Imperfect” ECG signals thus offer biologically plausible and trustworthy inputs for standalone or ensemble artificial intelligence models, enabling transparent, mode-aware decision support in real-time fetal monitoring. Health sciences/Cardiology Health sciences/Health care Health sciences/Medical research Figures Figure 1 Figure 2 Figure 3 Introduction Perinatal asphyxia, often attributed to intrapartum compromise of gas exchange, remains a significant contributor to neonatal illness and death—though precise incidence and definitions vary globally. Estimates suggest up to one-quarter of neonatal deaths and a large proportion of stillbirths in the third trimester are related to this condition. 1-3 Reported rates of intrapartum-related neonatal encephalopathy vary, with estimates ranging from 1–3 per 1,000 in high-income countries to significantly higher in low-resource settings. 3,4 Despite widespread use of cardiotocography (CTG) for fetal surveillance, outcomes have not improved, while cesarean rates and medicolegal costs have surged. 2,5,6 Machine learning models for CTG interpretation have shown limited impact, often hindered by small datasets, inadequate adjustment for confounding, and exclusion of signal segments deemed ‘noisy’ or ‘artifact’—segments that may still contain clinically relevant patterns. 7-9 Excessive preprocessing may obscure associations with asphyxia risk. The rarity of asphyxia further limits statistical power, and a lack of causal discovery risks misattributing predictive importance to spurious features. 7,10,11 Fetal heart rate (FHR) is typically acquired by CTG monitors at ~4 Hz using either Doppler ultrasound or fetal scalp electrode (electrocardiography, ECG). The former captures mechanical motion while the latter directly records cardiac electrical activity, generally offering greater signal fidelity during active labor. Mode selection is subjective by clinician—ECG use generally follows failed ultrasound acquisition during labor. Signal dropout is stored as zeros in backend systems and displayed as trace gaps. Although proprietary algorithms enhance visual readability through autocorrelation and filtering, this may conceal clinically important data loss. 7,12 Dropout differs by acquisition mode: ultrasound is more prone to signal loss due to maternal or fetal factors, while ECG—despite directly capturing cardiac activity—still exhibits dropout, a counterintuitive finding that may have a relationship with underlying fetal compromise. FDA reports have cited signal quality failures in adverse outcomes. 8 Across disciplines, the challenge of distinguishing signal loss due to technical limitations versus physiological compromise is shared in many biomedical domains—from ECG and EEG to wearables and implantables. 13 Our findings contribute to this growing body of work, demonstrating that signal dropout can act as a latent biomarker, especially when stratified by acquisition context. This work also speaks to broader trends in AI-driven health monitoring: the transition from pattern recognition to causal, real-time inference from raw and imperfect signals. 11 As CTG becomes increasingly digitized, signal quality itself may hold prognostic value. We applied causal machine learning to assess whether mode-specific FHR dropout predicts perinatal asphyxia. By treating signal dropout as an informative exposure rather than dismissing it as nuisance noise, we aim to support the development of real-time, mode-aware decision-support tools in fetal monitoring. By uncovering underlying causal drivers using causal machine learning—including those embedded in signal quality—we seek to inform more robust and generalizable prediction models for real-world clinical deployment. MATERIALS AND METHODS Study Design and Data Sources This multicenter cohort study included 52,757 laboring women at ≥36 weeks’ gestation across three international datasets (Figure 1). The training cohort consisted of 36,792 births from Mercy Hospital for Women, Melbourne (2010–2021).This is a tertiary hospoital. Inclusion required ≥15 minutes of in-labor CTG data within the final 60 minutes before birth, regardless of labor stage or delivery mode, to ensure inclusion of both vaginal and cesarean births. This aligns with standards in computational fetal monitoring studies. 14,15 Multiple gestations and congenital anomalies were excluded. Active labor was confirmed by midwifery and obstetric annotations. Data Extraction and Preprocessing CTG signals were extracted from Philips IntelliSpace™ (Philips, Amsterdam) with technical support from Philips. Traces were linked to clinical data from the Birthing Outcomes System (BOS), hospital morbidity records, and paper charts as needed 7 . All data were stored securely using the University of Melbourne’s MediaFlux™ platform, with analysis conducted on the Spartan high-performance computing cluster. External Validation Datasets Validation was performed on two cohorts. The first comprised 552 cases from a European dataset -the Czech CTU-UHB database (2010–2012), an open-access European dataset with linked clinical metadata. 16 The second included 15,413 births from Werribee Mercy Hospital (2010–2021),a community hospital in Australia. The data was processed identically to the training dataset. Both required ≥15 minutes of final-hour labor CTG data and excluded multiple gestations and congenital anomalies. Quality Control CTG recordings with >90% signal dropout were excluded due to their clinical unreliability. Exposure Classification The primary exposure was a high level of missing valid fetal heart rate (FHR) signal, specifically of the signal dropout type—defined as the cumulative proportion of zero-valued data points (missing valid biosignal)recorded during active signal acquisition in a live fetus. In Philips systems, these dropouts appear as visual gaps in the CTG trace; however, proprietary signal-processing algorithms (e.g., autocorrelation-based smoothing) may reduce their visibility. Dropout was calculated at a sampling frequency of 4 Hz over the final hour of labor. Based on Delphi consensus from a multidisciplinary team of clinicians, data scientists, and signal-processing experts, high dropout was predefined as >30% signal loss. 7 Maternal–FHR coincidence was not assessed in this analysis. Participants were categorized into four exposure groups based on monitoring mode (ultrasound vs. ECG) and dropout level (low vs. high): Low dropout in ultrasound mode (reference) High dropout in ultrasound mode Low dropout in ECG mode High dropout in ECG mode Mode classification was based on the presence and duration of ECG signal. Patients were assigned to the ECG group if ≥10 minutes of ECG was recorded, minimizing misclassification from brief mode switches. Those with <10 minutes were classified as ultrasound. This approach enabled mode-specific dropout analysis while reducing bias from clinician-initiated switching. Outcome Measures The primary outcome was perinatal asphyxia, defined using a composite developed by multidisciplinary Delphi consensus and aligned with international standards. Criteria included one or more of: resuscitation at 10 minutes, therapeutic hypothermia, Apgar score ≤6 at 10 minutes or ≤4 at 5 minutes, cord pH <7.05, base excess <–12 mmol/L, NICU admission, or clinically (including seizures) /imaging-confirmed hypoxic–ischaemic encephalopathy (HIE). Causal Machine Learning Methods Confounder adjustment was guided by directed acyclic graphs (DAGs) (see Supplement). A doubly robust framework combining propensity score modeling and outcome regression was used to assess the causal effect of mode-specific dropout. All confounders used in causal models are listed in supplementary material and include maternal demographics, obstetric risk factors, labor characteristics, and mode of delivery. Propensity scores were estimated via multinomial logistic regression using maternal, fetal, and intrapartum variables. Inverse probability weights (IPWs), stabilized by clipping between 0.01 and 0.99, were used in weighted logistic regression. Covariate balance was assessed using standardized mean differences. HC3 robust standard errors accounted for heteroskedasticity. Sensitivity analyses were conducted on both complete case and imputed datasets. Targeted Maximum Likelihood Estimation (TMLE) refined causal estimates for ECG High Dropout vs. Ultrasound Low Dropout. TMLE combined logistic regression and gradient boosting for exposure and outcome modeling. Ridge regularization was applied for logistic regression; the gradient boosting model used a learning rate of 0.1, max depth 3, and 100 estimators. Categorical variables were one-hot encoded; continuous variables (e.g., BMI) were log-transformed and standardized. Additional engineered features, such as labor duration ratios, were included. TMLE used a clever covariate (H) for targeted updates, and influence-curve–based standard errors (via the zepid package) for confidence intervals. No additional fetal heart rate features (e.g., decelerations or variability) were modeled directly to isolate the effect of dropout. Model Validation and Sensitivity Analysis Robustness was evaluated across dropout thresholds using likelihood ratio tests (LRTs) and E-values for unmeasured confounding. Models were stratified by labor stage and tested on a 20% stratified holdout set from the training cohort. External validation was conducted using both the Czech and Werribee datasets. Survival Analysis Time-to-event analysis assessed asphyxia risk from second-stage labor onset to event occurrence. "Survival" denoted remaining free from asphyxia. Exposure was categorized as a four-level composite of monitoring mode and dropout severity. Kaplan–Meier curves were compared using log-rank tests. A weighted Cox proportional hazards model, adjusted using stabilized IPWs derived from a multinomial logistic regression, was used to account for confounding. Model discrimination was assessed using Harrell’s concordance index (C-index). Sample size calculation Consistent with our prior study on CTG signal quality and perinatal outcomes—conducted on a related but differently grouped cohort—the current analysis builds on a pre-established dataset, including one additional large dataset, for which formal sample size justification had previously been undertaken 7 . Given the exploratory nature of this causal inference study and the rarity of perinatal asphyxia, we analyzed the full cohort (n = 52,757) to maximize statistical power for effect estimation, threshold modeling, and validation across independent populations. Data Processing and Statistical Analysis All analyses were performed using Python 3.6(including lifelines v0.30.0). and Stata 18.5. CTG data were extracted via Philips IntelliSpace™ and managed on Spartan and MediaFlux™ infrastructures. The detailed statistical analysis plan is enclosed in the supplementary material. Handling Missing Clinical Data Missing clinical variables were imputed using chained equations. Sensitivity analyses compared complete-case and imputed models. Maternal BMI had the highest missingness (6.32%, 95% CI: 6.06–6.66%). Ethical Approval and Reporting Standards The study was approved by the institutional Human Research Ethics Committee (HREC #R2020-077). It adhered to a pre-specified statistical analysis plan and followed STROBE 17 and TRIPOD+AI guidelines(see checklists in Supplementary Materials). 18 RESULTS Of 36,792 laboring women screened in the training dataset, 32,242 met the inclusion criteria. There were 860 cases of perinatal asphyxia (2.67%, 95% CI: 2.49–2.85%) (Figure1). Table 1 shows baseline sample characteristics. Sample characteristics of the validation datasets are included in supplementary tables 1 and 2. Dropout patterns significantly varied by monitoring modality. Most cases were in the Ultrasound Low Dropout group (Group 0, 54.4%, 95% CI: 53.8%–54.9%), which had the lowest asphyxia rate (2.2%, 95% CI: 2.0%–2.4%). In contrast, the ECG High Dropout group (Group 3, 1.18%, 95% CI: 1.07%–1.31%) had the highest asphyxia rate (5.2%, 95% CI: 3.4%–7.9%)(p < 0.0001) (Figure 1). Maternal and neonatal characteristics differed across groups. Maternal BMI was highest in ECG groups, with a median of 25.0 (IQR: 22.0–29.0) compared to 23.0 (IQR: 21.0–27.0) in the US Low Dropout group (p < 0.0001). Gestational age was similar across groups (p = 0.09), but baby weight was significantly lower in ECG High Dropout (3280g, IQR: 2970–3610) compared to US Low Dropout (3410g, IQR: 3110–3710, p < 0.0001). Parity distributions also varied, with nulliparous women being most common in the ECG Low Dropout group (64.7%) versus 47.5% in US Low Dropout (p < 0.0001). Figure 2 shows representative modal devices and representative outputs. Causal Inference Analysis Logistic regression with inverse probability weighting (IPW) was used to estimate the association between signal dropout and perinatal asphyxia, using Ultrasound Low Dropout as the reference group (Figure 3a). The highest risk of asphyxia was observed in the ECG High Dropout group (OR = 3.14, 95% CI: 2.20–4.48, p < 0.001). The ECG Low Dropout group also showed significantly increased risk (OR = 1.51, 95% CI: 1.32–1.71, p < 0.001). The Ultrasound High Dropout group showed a non-significant risk of asphyxia (OR = 1.25, 95% CI: 0.99–1.58, p = 0.059) (table 2, Figure 3a). E-value analysis assessed robustness to unmeasured confounding. The ECG High Dropout group showed the greatest robustness (E-value = 5.73; lower bound = 3.82). ECG Low Dropout showed moderate robustness (E-value = 2.39), while Ultrasound High Dropout , which was not statistically significant, showed limited robustness (E-value = 1.81). A likelihood ratio test (LRT) confirmed that the inclusion of the 'dropout by mode' classification significantly improved model fitcompared to a model without it (χ² (1) = 25.58, p < 0.0001) supporting the relevance of mode-specific dropout classification in explaining variation in asphyxia risk. Causal Machine Learning Analysis Targeted Maximum Likelihood Estimation (TMLE) refined causal estimates for the two key groups (US Low Dropout as reference and ECG High Dropout with highest asphyxia risk) by integrating machine learning–based propensity score modeling (Gradient Boosting Classifier) and outcome prediction (Gradient Boosting Regressor) to address high-dimensional, nonlinear relationships. TMLE confirmed increased asphyxia risk in ECG High Dropout (adjusted OR = 2.08, 95% CI: 1.21–3.58) relative to US Low Dropout (Table 2). In crude analyses, ECG High Dropout was associated with a 5.0% absolute risk increase—one additional case of asphyxia for every 20 laboring women monitored. After TMLE adjustment, the absolute risk increase was 2.27%, or one extra case per 44 labor episodes (95% CI: 28–83). We applied TMLE-derived individual risk estimates to assess fetal risk reclassification in the training set. Clinically, cases were considered low risk if labor progressed without operative delivery, and high risk if operative birth (caesarean, forceps, or vacuum) occurred due to suspected fetal distress. TMLE reclassified 9.29% of cases managed as low-risk into a high-risk category based on elevated model-predicted asphyxia risk. Reclassification was significantly more frequent among ECG-monitored cases, suggesting dropout-informed causal modeling may identify high-risk fetuses missed by conventional assessment. TMLE analyses across dropout thresholds showed progressively increasing odds of asphyxia. Dropout <15% was not significantly associated with asphyxia, while increasing dropout thresholds from ≥20% showed a dose-response relationship with rising risk: OR = 1.56 (95% CI: 1.16–2.09, p < 0.001) at 20%, increasing to OR = 2.08 (95% CI: 1.21–3.58, p = 0.01) at 30% (Figure 3a). SHAP analysis in Supplementary Figure 3 illustrates features driving ECG High Dropout and asphyxia risk. Survival Analysis (Figure 3b) Kaplan–Meier estimates stratified by monitoring modality and signal dropout showed the highest asphyxia-free survival in the Ultrasound Low Dropout group, and the steepest decline in the ECG High Dropout group. Differences were statistically significant (log-rank χ² = 43.6, p < 0.0001). To adjust for confounding, we applied an inverse probability–weighted Cox proportional hazards model using stabilized weights from a multinomial treatment model. In the fully adjusted model, the hazard of asphyxia was significantly higher in ECG High Dropout versus US Low Dropout (HR = 4.76, 95% CI: 2.98–7.62, p < 0.0001). Elevated hazards were also observed for Ultrasound High Dropout (HR = 1.90, 95% CI: 1.39–2.60, p < 0.0001) and ECG Low Dropout (HR = 1.42, 95% CI: 1.20–1.69, p < 0.0001) (Supplementary Table 3). Harrell’s C-index was 0.71, indicating moderate predictive accuracy. Figure 3b displays Kaplan–Meier and IPW-adjusted Cox curves by monitoring modality and dropout level. External Validation External validation using the Czech CTU-UHB and Werribee datasets demonstrated similar trends to the model generation cohort(table 3) Validation on Czech dataset Causal inference analysis using inverse probability weighting (IPW) demonstrated that the ECG High Dropout group was significantly associated with an increased risk of perinatal asphyxia (OR = 4.02, 95% CI: 1.41–11.44, p = 0.009) compared to the Ultrasound Low Dropout reference group. The ECG Low Dropout group showed a paradoxical reduction in asphyxia risk (OR = 0.24, 95% CI: 0.06–0.96, p = 0.043), which may reflect differing case mix, small sample size, or differences in mode-switching thresholds between centers. The Ultrasound High Dropout group showed a non-significant trend toward increased risk (OR = 1.90, 95% CI: 0.94–3.82, p = 0.073). Causal machine learning analysis using targeted maximum likelihood estimation (TMLE) produced consistent findings, estimating an asphyxia odds ratio of 2.83 (95% CI: 2.13–16.36) for the ECG High Dropout group compared to the reference Ultrasound Low Dropout . Validation on Werribee dataset Causal inference analysis using inverse probability weighting (IPW) regression model showed that the ECG High Dropout group was significantly associated with increased asphyxia risk (OR = 2.22, 95% CI: 1.61–3.05, p < 0.001), followed by the ECG Low Dropout group (OR = 1.57, 95% CI: 1.31–1.89, p < 0.001). The Ultrasound High Dropout group was not significantly associated with asphyxia (OR = 0.69, 95% CI: 0.33–1.46, p = 0.33), using Ultrasound Low Dropout as the reference. Causal machine learning using targeted maximum likelihood estimation (TMLE) showed that the ECG High Dropout group showed a significant increase in asphyxia risk (OR = 1.91, 95% CI: 1.11–3.28) compared to the reference Ultrasound Low Dropout . DISCUSSION Principal Findings In this multicohort study of over 50,000 births, we show that mode-specific FHR signal dropout—typically overlooked in CTG interpretation—is strongly and causally associated with perinatal asphyxia, particularly in ECG-based monitoring. ECG High Dropout carried the highest risk (IPW OR = 3.14; 95% CI: 2.20–4.48), followed by ECG Low Dropout (OR = 1.51; 95% CI: 1.32–1.71). Ultrasound High Dropout was not significant, reinforcing the hypothesis that dropout in ECG—typically used in advanced labor—may more directly reflect underlying physiological compromise rather than technical factors. TMLE confirmed a dose-response relationship, with dropout ≥30% yielding an OR of 2.08 and an absolute risk increase of one asphyxia case per 44 women. TMLE-based risk profiling reclassified nearly 10% of clinically low-risk cases to high risk. Findings were consistent across external cohorts (Czech: IPW OR = 4.02, TMLE OR = 2.83; Werribee: IPW OR = 2.22, TMLE OR = 1.91), reinforcing the prognostic value of ECG dropout. Results in the Context of What is Known Signal dropout in ECG-mode CTG is a strong, clinically meaningful predictor of asphyxia—challenging the assumption that it reflects only technical or user error. Mode selection is subjective, and dropout may get hidden due to interface design. Our findings support making dropout metrics visible. Rather than default mode switching, decisions should be guided by dropout severity: US-mode dropout >30% may warrant switching to ECG, but persistent high dropout in ECG-mode indicates tripled asphyxia risk and should prompt closer surveillance Clinical Implications A major strength of this study is the use of causal methods—IPW and TMLE—to estimate the effect of mode-specific dropout. Dropout is non-random and influenced by clinical factors; failure to adjust for this can bias observational analyses. Our approach improved covariate balance and model fit (χ²(1) = 16.2, p = 0.00006), with E-values (ECG High Dropout: 5.89) supporting robustness to unmeasured confounding. Unlike traditional machine learning, which may overfit and overlook causal pathways, causal ML methods like TMLE explicitly model exposure–outcome relationships to enhance generalizability. 11,19-21 We used DAGs to select confounders and validated findings across two independent cohorts. Prior models, such as McCoy et al., excluded high-artifact CTGs (>30% dropout), discarding over one-third of cases. 22 While their deep learning model achieved high AUROC (0.85 for pH <7.05), it ignored dropout as a signal. Moreover, pH <7 alone does not reliably stratify neonatal outcomes, as most infants with pH between 6.9 and 7 survive without neurodevelopmental impairment. 23 In contrast, we treat signal dropout as a meaningful exposure with prognostic value and applied a more holistic, clinically accepted composite outcome aligned with established definitions of neurodevelopmental risk. 5,24,25 While our model focused on signal dropout, it did not incorporate contemporaneous CTG pattern interpretation (e.g., decelerations), which may offer complementary prognostic information in future models. Research Implications By integrating mode-specific dropout into causal ML, we demonstrate its utility in identifying high-risk cases missed by conventional approaches. This supports real-time monitoring tools that adapt to dropout severity, aligning with current recommendations for actionable, causally grounded AI in healthcare. 11,19-21 Beyond obstetrics, our findings resonate with challenges in other domains that rely on physiological time series—such as ECG, EEG, and wearable biosensors—where distinguishing noise from latent risk is critical for advancing real-time, interpretable AI applications. Strengths and Limitations Our study leverages one of the largest digitized CTG datasets, with validation across multiple cohorts, enhancing generalizability. The integration of causal inference and machine learning improves robustness. 12,19,20,26,27 A limitation is the lack of data on real-time clinician-driven mode-switching decisions. However, consistent validation trends across Czech CTU-UHB and Werribee datasets reinforce the observed associations. Implementation in other settings may face logistical challenges, but prospective validation is essential to assess whether integrating dropout monitoring improves neonatal outcomes. We acknowledge the lack of data on real-time clinician-driven mode-switching or intrauterine resuscitation interventions, which may confound the relationship between dropout and outcome While our framework enhances risk estimation, clinical deployment would require real-time implementation within a comprehensive decision-support model. Future extensions may adapt approaches such as the AI-ECG sex discordance score, which identifies subtle, clinician-invisible signal shifts—analogous to dropout thresholds in our study. 28 However, developers must also address known demographic disparities in signal-based AI, as shown in recent ECG heart failure models, where performance biases were especially pronounced among young Black women. 29 This study also contributes to emerging efforts across disciplines to better utilize ‘imperfect’ or artifact-prone physiological data streams—turning what was once discarded into features that enhance clinical relevance and model transparency. Conclusion Mode-specific FHR dropout—particularly in ECG-mode CTG—is a clinically significant predictor of perinatal asphyxia. Rather than being excluded or ignored, signal quality should be incorporated into predictive modeling. Integrating signal metrics into ensemble models may enhance fetal surveillance and should be prioritized in future validation efforts. Declarations Disclosure/Conflict of interest: None of the authors have a direct conflict of interest. FB and EK are directors and shareholders of Kali Healthcare, a company that is commercialising a wearable fetal monitoring device. MP is a shareholder of Kali Healthcare. Rest of the authors have nothing to disclose. Funding FB is supported by a Dame Kate Campbell Fellowship, a University of Melbourne Fellowship, and an National Health and Medical Research Council (NHMRC) Ideas Grant. DK receives postgraduate scholarship support from the NHMRC.Additional funding was provided by the Norman Beischer Medical Research Foundation. LM receives University of Melbourne Graeme Clark Institute and Melbourne Research scholarships. The authors were not precluded from accessing data in the study, and they accept responsibility to submit for publication. Acknowledgments: This research was supported by The University of Melbourne’s Research Computing Services and the Petascale Campus Initiative. Funding: FB is supported by a Dame Kate Campbell Fellowship, a University of Melbourne Fellowship, and an National Health and Medical Research Council (NHMRC) Ideas Grant. DK receives postgraduate scholarship support from the NHMRC.Additional funding was provided by the Norman Beischer Medical Research Foundation. LM receives University of Melbourne Graeme Clark Institute and Melbourne Research scholarships. The authors were not precluded from accessing data in the study, and they accept responsibility to submit for publication. Author contributions: DK: Conceptualization, Methodology, Investigation, Data curation, Writing – review & editing Visualization, Funding acquisition, Writing – original draft FB: Conceptualization, Methodology, Investigation, Funding acquisition, Project administration, Supervision, Writing – original draft EM: Methodology, Visualization, Project administration, Supervision, Writing – review & editing LM: Investigation, Data curation, Writing – review & editing Writing – review & editing EK: Supervision, Data curation, Writing – review & editing Writing – review & editing MP: Supervision, Writing – review & editing Competing interests: None of the authors have a direct conflict of interest. FB and EK are directors and shareholders of Kali Healthcare, a company that is commercialising a wearable fetal monitoring device. MP is a shareholder of Kali Healthcare. Rest of the authors have nothing to disclose. Data and materials availability: All individual-level data of the included cohort and modeling codes can be shared upon reasonable request to the corresponding author and completion of data transfer agreement forms. References Mota-Rojas, D. , et al. Pathophysiology of perinatal asphyxia in humans and animal models. Biomedicines 10 , 347 (2022). Resolution, N.H.S. Annual Report and Accounts 2019/20. (NHS Resolution, UK, 2020). Schwartz, N. & Young, B.K. Intrapartum fetal monitoring today. (2006). Schiermeier, S. , et al. Sensitivity and specificity of intrapartum computerised FIGO criteria for cardiotocography and fetal scalp pH during labour: multicentre, observational study. Bjog 115 , 1557-1563 (2008). Brocklehurst, P. , et al. Computerised interpretation of fetal heart rate during labour (INFANT): a randomised controlled trial. The Lancet 389 , 1719-1729 (2017). Ayres-de-Campos, D., Spong, C.Y. & Chandraharan, E. FIGO consensus guidelines on intrapartum fetal monitoring: Cardiotocography. International Journal of Gynecology & Obstetrics 131 , 13-24 (2015). Karmakar, D. , et al. Impact of missing electronic fetal monitoring signals on perinatal asphyxia: a multicohort analysis. NPJ Digit Med 8 , 233 (2025). Kiely, D.J., Oppenheimer, L.W. & Dornan, J.C. Unrecognized maternal heart rate artefact in cases of perinatal mortality reported to the United States Food and Drug Administration from 2009 to 2019: a critical patient safety issue. BMC Pregnancy and Childbirth 19 , 1-10 (2019). Signorini, M.G. & Magenes, G. Advanced signal processing techniques for CTG analysis. in XIV Mediterranean Conference on Medical and Biological Engineering and Computing 2016: MEDICON 2016, March 31st-April 2nd 2016, Paphos, Cyprus 1205-1210 (Springer, 2016). O’Sullivan, M.E. , et al. Challenges of developing robust AI for intrapartum fetal heart rate monitoring. Frontiers in Artificial Intelligence 4 , 765210 (2021). Prosperi, M. , et al. Causal inference and counterfactual prediction in machine learning for actionable healthcare. Nature Machine Intelligence 2 , 369-375 (2020). Blakely, T., Lynch, J., Simons, K., Bentley, R. & Rose, S. Reflection on modern methods: when worlds collide—prediction, machine learning and causal inference. International Journal of Epidemiology 49 , 2058-2064 (2019). Clifford, G.D., Behar, J., Li, Q. & Rezek, I. Signal quality indices and data fusion for determining clinical acceptability of electrocardiograms. Physiol Meas 33 , 1419-1433 (2012). Mendis, L., Palaniswami, M., Brownfoot, F. & Keenan, E. Computerised cardiotocography analysis for the automated detection of fetal compromise during labour: a review. Bioengineering 10 , 1007 (2023). Petrozziello, A., Redman, C.W., Papageorghiou, A.T., Jordanov, I. & Georgieva, A. Multimodal convolutional neural networks to detect fetal compromise during labor and delivery. IEEE Access 7 , 112026-112036 (2019). Chudáček, V. , et al. Open access intrapartum CTG database. BMC pregnancy and childbirth 14 , 1-12 (2014). Von Elm, E. , et al. The Strengthening the Reporting of Observational Studies in Epidemiology (STROBE) statement: guidelines for reporting observational studies. The lancet 370 , 1453-1457 (2007). Collins, G.S. , et al. TRIPOD+ AI statement: updated guidance for reporting clinical prediction models that use regression or machine learning methods. bmj 385 (2024). Hernán, M.A. & Robins, J.M. Using Big Data to Emulate a Target Trial When a Randomized Trial Is Not Available. Am J Epidemiol 183 , 758-764 (2016). Hernán, M.A., Wang, W. & Leaf, D.E. Target trial emulation: a framework for causal inference from observational data. Jama 328 , 2446-2447 (2022). Nevo, D., Ogino, S. & Wang, M. Reflection on modern methods: causal inference considerations for heterogeneous disease etiology. International Journal of Epidemiology 50 , 1030-1037 (2021). McCoy, J.A. , et al. Intrapartum electronic fetal heart rate monitoring to predict acidemia at birth with the use of deep learning. American Journal of Obstetrics and Gynecology (2024). Lau, S.L. , et al. Neonatal outcome of infants with umbilical cord arterial pH less than 7. Acta Obstet Gynecol Scand 102 , 174-180 (2023). Lee, A.C. , et al. Intrapartum-related neonatal encephalopathy incidence and impairment at regional and global levels for 2010 with trends from 1990. Pediatric research 74 , 50-72 (2013). Razak, A. , et al. Early Neurodevelopmental Assessments for Predicting Long-Term Outcomes in Infants at High Risk of Cerebral Palsy. JAMA Network Open 7 , e2413550-e2413550 (2024). Luque‐Fernandez, M.A., Schomaker, M., Rachet, B. & Schnitzer, M.E. Targeted maximum likelihood estimation for a binary treatment: A tutorial. Statistics in medicine 37 , 2530-2546 (2018). Smith, M.J., Phillips, R.V., Luque-Fernandez, M.A. & Maringe, C. Application of targeted maximum likelihood estimation in public health and epidemiological studies: a systematic review. Annals of epidemiology 86 , 34-48. e28 (2023). Sau, A. , et al. Artificial intelligence-enhanced electrocardiography for the identification of a sex-related cardiovascular risk continuum: a retrospective cohort study. The Lancet Digital Health 7 , e184-e194 (2025). Kaur, D. , et al. Race, sex, and age disparities in the performance of ECG deep learning models predicting heart failure. Circulation: Heart Failure 17 , e010879 (2024). Tables Table 1: Baseline characteristics by mode specific dropout levels US mode, low dropout US mode, high dropout ECG mode, low dropout ECG mode, high dropout P-value Number of records, % proportion, (95%CI) 17533, (54.38,53.83-54.92) 2363, (7.33,7.05-7.62) 11964, (37.11,36.58-37.64) 382(1.18,1,07-1.31) 0.00 Proportion of CTG signal by US mode 1.00 (1.00-1.00) 1.00 (1.00-1.00) 0.04 (0.00-0.35) 0.70 (0.57-0.83) 0.00 Maternal age 32.00 (29.00-35.00) 32.00 (29.00-35.00) 32.00 (29.00-35.00) 32.50 (29.00-36.00) 0.00 Gestational age 39.40 (38.50-40.30) 39.40 (38.50-40.30) 39.50 (38.50-40.40) 39.40 (38.40-40.30) 0.09 Maternal BMI 23.00 (21.00-27.00) 24.00 (21.00-28.00) 25.00 (22.00-29.00) 25.00 (22.00-29.00) <0.0001 Baby Weight(grams) 3410.00 (3110.00-3710.00) 3380.00 (3070.00-3690.00) 3350.00 (3040.00-3660.00) 3280.00 (2970.00-3610.00) <0.0001 Parity Nulliparous 47.5% (0.468 - 0.483) 35.3% (0.334 - 0.373) 64.7% (0.638 - 0.655) 51.0% (0.460 - 0.560) <0.0001 Para 1 33.8% (0.331 - 0.345) 46.0% (0.440 - 0.481) 24.3% (0.236 - 0.251) 35.6% (0.310 - 0.405) Para 2 12.9% (0.124 - 0.134) 13.2% (0.119 - 0.147) 7.7% (0.073 - 0.082) 9.4% (0.069 - 0.128) Para 3 or higher 5.7% (0.054 - 0.061) 5.4% (0.045 - 0.064) 3.2% (0.029 - 0.036) 3.9% (0.024 - 0.064) Baby gender Female/indeterminate 49.6% (0.489 - 0.503) 50.0% (0.480 - 0.520) 47.6% (0.467 - 0.484) 47.9% (0.429 - 0.529) 0.00 Male 50.4% (0.497 - 0.511) 50.0% (0.480 - 0.520) 52.4% (0.516 - 0.533) 52.1% (0.471 - 0.571) Primary Obstetrics Diagnosis Low risk pregnancy 44.0% (0.433 - 0.448) 60.0% (0.580 - 0.620) 35.3% (0.344 - 0.361) 48.4% (0.435 - 0.534) <0.0001 Fetal compromise † 19.9% (0.193 - 0.205) 12.0% (0.107 - 0.133) 25.3% (0.246 - 0.261) 16.0% (0.126 - 0.200) Miscellaneous fetal conditions †† 3.2% (0.029 - 0.034) 1.7% (0.013 - 0.023) 3.7% (0.033 - 0.040) 1.8% (0.009 - 0.037) Maternal medical conditions § 16.9% (0.164 - 0.175) 14.3% (0.129 - 0.157) 19.2% (0.185 - 0.200) 19.1% (0.155 - 0.234) Pre labor rupture of membranes 12.1% (0.117 - 0.126) 9.2% (0.081 - 0.104) 13.0% (0.124 - 0.136) 11.5% (0.087 - 0.151) Onset of labor Induced 49.2% (0.485 - 0.500) 29.2% (0.274 - 0.311) 60.8% (0.600 - 0.617) 43.5% (0.386 - 0.485) <0.0001 Spontaneous 50.8% (0.500 - 0.515) 70.8% (0.689 - 0.726) 39.2% (0.383 - 0.400) 56.5% (0.515 - 0.614) Fetal presentation Vertex 97.5% (0.972 - 0.977) 97.4% (0.967 - 0.980) 98.4% (0.982 - 0.986) 97.9% (0.959 - 0.989) <0.0001 Others 2.5% (0.023 - 0.028) 2.6% (0.020 - 0.033) 1.6% (0.014 - 0.018) 2.1% (0.011 - 0.041) Fetal Position Non-Vertex 2.5% (0.023 - 0.028) 2.6% (0.020 - 0.033) 1.6% (0.014 - 0.018) 2.1% (0.011 - 0.041) <0.0001 Vertex, occiput anterior 53.5% (0.528 - 0.543) 46.4% (0.444 - 0.484) 60.6% (0.597 - 0.614) 52.4% (0.473 - 0.573) Vertex, not occiput anterior 43.9% (0.432 - 0.447) 51.0% (0.490 - 0.530) 37.8% (0.370 - 0.387) 45.5% (0.406 - 0.506) Maternal position Dorsal/lithotomy/semi recumbent 90.1% (0.896 - 0.905) 79.6% (0.779 - 0.811) 92.7% (0.922 - 0.931) 88.7% (0.852 - 0.915) <0.0001 Birth stool/squatting/Standing/Water 1.3% (0.012 - 0.015) 2.5% (0.019 - 0.032) 0.7% (0.006 - 0.009) 0.8% (0.003 - 0.023) All fours/kneeling 5.2% (0.049 - 0.056) 12.1% (0.108 - 0.135) 4.1% (0.038 - 0.045) 6.5% (0.045 - 0.095) Lateral 3.4% (0.031 - 0.036) 5.9% (0.050 - 0.069) 2.5% (0.023 - 0.028) 3.9% (0.024 - 0.064) Regional Anesthesia use in labor 42.9% (0.422 - 0.437) 13.2% (0.119 - 0.146) 51.9% (0.510 - 0.528) 24.6% (0.206 - 0.292) Duration of labor Total 4.97 (2.70-8.35) 3.80 (2.22-6.23) 5.47 (3.20-8.53) 3.62 (2.02-6.40) <0.0001 First stage 4.17 (2.25-7.25) 3.33 (1.88-5.50) 4.58 (2.58-7.50) 3.12 (1.67-5.66) Second stage 0.47 (0.18-1.20) 0.35 (0.18-0.68) 0.50 (0.17-1.22) 0.33 (0.17-0.58) Intrapartum passage of meconium 12.9% (0.125 - 0.135) 11.2% (0.100 - 0.125) 14.5% (0.139 - 0.152) 17.3% (0.138 - 0.214) <0.0001 Sentinel events Cord accident 16.5% (0.160 - 0.171) 17.5% (0.160 - 0.191) 15.4% (0.148 - 0.160) 16.5% (0.131 - 0.205) 0.07 Shoulder dystocia/abruption/failed instrumental 1.2% (0.011 - 0.014) 1.4% (0.010 - 0.019) 1.2% (0.010 - 0.014) 1.8% (0.009 - 0.037) Mode of birth Vaginal 70.4% (0.697 - 0.711) 90.1% (0.888 - 0.912) 52.1% (0.512 - 0.530) 67.3% (0.624 - 0.718) 0.00 Forceps assisted 10.9% (0.105 - 0.114) 2.5% (0.019 - 0.032) 16.5% (0.159 - 0.172) 8.9% (0.064 - 0.122) Vacuum assisted 8.4% (0.080 - 0.088) 4.4% (0.036 - 0.053) 13.8% (0.132 - 0.144) 17.5% (0.141 - 0.217) CS 10.3% (0.099 - 0.108) 3.0% (0.024 - 0.038) 17.6% (0.169 - 0.182) 6.3% (0.043 - 0.092) Incidence of asphyxia* (number of records, % proportion,95%CI) 383, 2.2 (1.98 - 2.41) 57, 2.4% (1.87 - 3.11) 400, 3.3% (3.03 - 3.68) 20, 5.2 (3.41 - 7.94) <0.0001 Data are median (IQR)(as not normally distributed), n (%,95% Confidence Interval). Only BMI had missing data as described * By composite definition; † e.g. Abruption, post-term; § Including hypertensive disease in pregnancy; †† e.g. macrosomia, polyhydramnios, unstable lie This table summarizes the clinical, obstetric, and neonatal characteristics of training cohort women stratified by fetal heart rate monitoring modality (ultrasound or ECG) and signal dropout level (>30% or ≤30%). Data are presented as medians (IQR) or proportions with 95% confidence intervals. Differences across groups were assessed using appropriate statistical tests. Table 2: Causal Effects of Mode-Specific Signal Dropout on Perinatal Asphyxia in Training and Validation Cohorts Dropout Group Model Adjusted Odds Ratio (aOR) a 95% CI p-value Training dataset ECG Low Dropout IPW 1.51 1.32-1.71 <0.001 US High Dropout 1.25 0.99-1.58 0.06 ECG High Dropout 3.14 2.20-4.48 <0.001 ECG High Dropout at threshold 5% TMLE 1.22 0.84-1.78 0.29 ECG High Dropout at threshold 10% 1.19 0.94-1.51 0.15 ECG High Dropout at threshold 15% 1.21 0.94-1.55 0.14 ECG High Dropout at threshold 20% 1.56 1.16-2.09 0.00 ECG High Dropout at threshold 25% 1.74 1.18-2.56 0.00 ECG High Dropout at threshold 30% 2.08 1.21-3.58 0.01 External Validation Dataset Czech dataset ECG High Dropout IPW b 4.02 1.41-11.44 0.01 ECG Low Dropout 0.24 0.06-0.96 0.04 US High Dropout 1.90 0.94-3.82 0.07 ECG High Dropout at threshold 30% TMLE 2.83 2.13-16.36 0.01 Australian dataset 2 (Werribee) ECG High Dropout IPW b 2.22 1.61-3.05 <0.001 ECG Low Dropout 1.57 1.31-1.89 <0.001 US High Dropout 0.69 0.33-1.46 0.33 ECG High Dropout at threshold 30% TMLE 1.91 1.11-3.28 0.02 a All comparisons are made against the Ultrasound Low Dropout group, which serves as the reference category in both IPW and TMLE models. b All IPW modelling was on a 30% dropout cutoff Adjusted odds ratios (aORs) from inverse probability–weighted logistic regression and dose-dependent targeted maximum likelihood estimation (TMLE) analyses show the association between mode-specific signal dropout and perinatal asphyxia in training and validation cohorts. Additional Declarations Competing interest reported. None of the authors have a direct conflict of interest. FB and EK are directors and shareholders of Kali Healthcare, a company that is commercialising a wearable fetal monitoring device. MP is a shareholder of Kali Healthcare. Rest of the authors have nothing to disclose. Supplementary Files SupplementaryMaterialmodesubmit.pdf Cite Share Download PDF Status: Posted Version 1 posted You are reading this latest preprint version Research Square lets you share your work early, gain feedback from the community, and start making changes to your manuscript prior to peer review in a journal. As a division of Research Square Company, we’re committed to making research communication faster, fairer, and more useful. We do this by developing innovative software and high quality services for the global research community. Our growing team is made up of researchers and industry professionals working together to solve the most critical problems facing scientific publishing. Also discoverable on Platform About Our Team In Review Editorial Policies Advisory Board Help Center Resources Author Services Accessibility API Access RSS feed Manage Cookie Preferences © Research Square 2026 | ISSN 2693-5015 (online) Privacy Policy Terms of Service Do Not Sell My Personal Information {"props":{"pageProps":{"initialData":{"identity":"rs-7122719","acceptedTermsAndConditions":true,"allowDirectSubmit":true,"archivedVersions":[],"articleType":"Article","associatedPublications":[],"authors":[{"id":496899710,"identity":"b3883866-a3c3-40e8-b4e3-5532143b0be5","order_by":0,"name":"Debjyoti Karmakar","email":"data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAAZAAAAAyAQMAAABI0h/eAAAABlBMVEX///8AAABVwtN+AAAACXBIWXMAAA7EAAAOxAGVKw4bAAAA60lEQVRIiWNgGAWjYDCCAzxAouAAhPMBiNnYidJiANHCOAOkhZkULcw8YJKADr7jvUc3fDC4I2dw++zBxza/tsnzMTMwfviYg1uL5JlzaTdnGDwzNjiXl2yc23fbsI2ZgVly5jbcWgxu5Jjd5jE4nLjhDI+ZdG7PbUagFjZmXnxa7r8xu/0HosX8t2XPbXvCWm7wmN1mgNrCzPDjdiJBLZJn8tJu9gD9InmGx1iyt+F2chszYzNev/AdP3vsxo+KO3J8Z3gMP/z4c9t2fnvzwQ8f8WhBBYxtYLKBWPUg8IcUxaNgFIyCUTBSAACg2lfo1Zs+WwAAAABJRU5ErkJggg==","orcid":"","institution":"University of Melbourne","correspondingAuthor":true,"prefix":"","firstName":"Debjyoti","middleName":"","lastName":"Karmakar","suffix":""},{"id":496899711,"identity":"79976efa-696f-4276-9ee9-d35991b0c0a8","order_by":1,"name":"Lochana Mendis","email":"","orcid":"","institution":"University of Melbourne","correspondingAuthor":false,"prefix":"","firstName":"Lochana","middleName":"","lastName":"Mendis","suffix":""},{"id":496899712,"identity":"29dbc417-b341-4161-9f17-41d4f3ebf1e8","order_by":2,"name":"Emerson Keenan","email":"","orcid":"","institution":"University of Melbourne","correspondingAuthor":false,"prefix":"","firstName":"Emerson","middleName":"","lastName":"Keenan","suffix":""},{"id":496899713,"identity":"d8a32061-c025-4b5e-ab80-e3600744d7f2","order_by":3,"name":"Marimuthu Palaniswami","email":"","orcid":"","institution":"University of Melbourne","correspondingAuthor":false,"prefix":"","firstName":"Marimuthu","middleName":"","lastName":"Palaniswami","suffix":""},{"id":496899714,"identity":"d1892ea5-a29f-4550-a967-887a42c16924","order_by":4,"name":"Enes Makalic","email":"","orcid":"","institution":"Monash University","correspondingAuthor":false,"prefix":"","firstName":"Enes","middleName":"","lastName":"Makalic","suffix":""},{"id":496899716,"identity":"2b834f55-90d4-4242-b75d-cfde9c70d9ec","order_by":5,"name":"Fiona Brownfoot","email":"","orcid":"","institution":"University of Melbourne","correspondingAuthor":false,"prefix":"","firstName":"Fiona","middleName":"","lastName":"Brownfoot","suffix":""}],"badges":[],"createdAt":"2025-07-14 15:23:32","currentVersionCode":1,"declarations":"","doi":"10.21203/rs.3.rs-7122719/v1","doiUrl":"https://doi.org/10.21203/rs.3.rs-7122719/v1","draftVersion":[],"editorialEvents":[],"editorialNote":"","failedWorkflow":false,"files":[{"id":88951568,"identity":"5ea311fa-3c9f-4239-9fec-c7e6e6a88421","added_by":"auto","created_at":"2025-08-13 05:53:58","extension":"jpg","order_by":1,"title":"Figure 1","display":"","copyAsset":false,"role":"figure","size":4729779,"visible":true,"origin":"","legend":"\u003cp\u003eMulticohort sample selection and analysis pipeline.\u003c/p\u003e\n\u003cp\u003e\u003cem\u003eFlow diagram showing inclusion criteria and sample selection across three cohorts used for model training and validation. The primary training dataset (Australia, n = 36,792), Czech validation set (n = 552), and a second Australian validation cohort (n = 15,413) are shown. All subjects were laboring at ≥36 weeks with interpretable CTG traces ≥15 minutes in the final 60 minutes before birth. Datasets were processed and harmonized using a unified preprocessing pipeline.\u003c/em\u003e\u003c/p\u003e","description":"","filename":"Fig1.tiff.jpg","url":"https://assets-eu.researchsquare.com/files/rs-7122719/v1/86bc61f34db91c6578418a40.jpg"},{"id":88951566,"identity":"b1ff9b98-0e8e-4392-b085-66570bf7768a","added_by":"auto","created_at":"2025-08-13 05:53:56","extension":"jpg","order_by":2,"title":"Figure 2","display":"","copyAsset":false,"role":"figure","size":3066734,"visible":true,"origin":"","legend":"\u003cp\u003eMode-specific cardiotocography(CTG) traces and dropout patterns.\u003c/p\u003e\n\u003cp\u003e\u003cem\u003e(a) CTG trace predominantly acquired in ultrasound mode(b) CTG trace predominantly in ECG mode. Representative CTG monitoring devices are shown alongside each trace (device models are illustrative and not necessarily identical to those used in the study). Traces reflect typical patterns observed in each exposure group, with paper speed standardized at 1 cm/min per Australian convention.\u003c/em\u003e\u003c/p\u003e","description":"","filename":"Fig2.tiff.jpg","url":"https://assets-eu.researchsquare.com/files/rs-7122719/v1/fcfd9fafb8f83b6ac4976832.jpg"},{"id":88951563,"identity":"ca1a9338-c9cc-483b-b4fa-8f672f2818ba","added_by":"auto","created_at":"2025-08-13 05:53:53","extension":"jpg","order_by":3,"title":"Figure 3","display":"","copyAsset":false,"role":"figure","size":159274,"visible":true,"origin":"","legend":"\u003cp\u003eRisk of Perinatal Asphyxia by Monitoring Modality and Signal Dropout: Causal Odds Ratios and Survival Analysis\u003c/p\u003e\n\u003cp\u003e\u003cem\u003ea)\u003c/em\u003e \u003cem\u003eForest plot summarizing the association between signal dropout and perinatal asphyxia across four mode-dropout exposure groups using causal inference (IPW) and causal machine learning (TMLE) in the primary Australian dataset. Odds ratios (ORs) with 95% confidence intervals are shown. The strongest association was observed in the ECG High Dropout group, with both IPW and TMLE methods indicating increased risk relative to the reference group (Ultrasound Low Dropout).\u003c/em\u003e\u003c/p\u003e\n\u003cp\u003e\u003cem\u003eb)\u003c/em\u003e \u003cem\u003eKaplan–Meier and IPW-adjusted Cox survival curves showing asphyxia-free survival across four exposure groups defined by monitoring modality and dropout level in the primary Australian dataset. (A) Unadjusted Kaplan–Meier curves demonstrate the steepest decline in the ECG High Dropout group. (B) IPW-adjusted Cox curves confirm persistently elevated hazards in this group, supporting its consistent association with higher asphyxia risk.\u003c/em\u003e\u003c/p\u003e","description":"","filename":"Fig3a.jpg","url":"https://assets-eu.researchsquare.com/files/rs-7122719/v1/292d5768383dd0008c86ea25.jpg"},{"id":90358470,"identity":"e145d342-be72-41a9-95b9-8a728d44b93c","added_by":"auto","created_at":"2025-09-01 23:31:28","extension":"pdf","order_by":0,"title":"","display":"","copyAsset":false,"role":"manuscript-pdf","size":9249457,"visible":true,"origin":"","legend":"","description":"","filename":"manuscript.pdf","url":"https://assets-eu.researchsquare.com/files/rs-7122719/v1/e6b53fef-e78d-485c-ba1e-66a44f3d5f5c.pdf"},{"id":88951567,"identity":"848e1487-d061-430c-a6f1-883c532cace7","added_by":"auto","created_at":"2025-08-13 05:53:57","extension":"pdf","order_by":0,"title":"","display":"","copyAsset":false,"role":"supplement","size":2459515,"visible":true,"origin":"","legend":"","description":"","filename":"SupplementaryMaterialmodesubmit.pdf","url":"https://assets-eu.researchsquare.com/files/rs-7122719/v1/4b7de9b28dae48e7205c0564.pdf"}],"financialInterests":"Competing interest reported. None of the authors have a direct conflict of interest. FB and EK are directors and shareholders of Kali Healthcare, a company that is commercialising a wearable fetal monitoring device. MP is a shareholder of Kali Healthcare. Rest of the authors have nothing to disclose.","formattedTitle":"Multimodal Causal Machine Learning for Fetal Asphyxia Risk Prediction from Imperfect Monitoring Signals ","fulltext":[{"header":"Introduction","content":"\u003cp\u003ePerinatal asphyxia, often attributed to intrapartum compromise of gas exchange, remains a significant contributor to neonatal illness and death\u0026mdash;though precise incidence and definitions vary globally. Estimates suggest up to one-quarter of neonatal deaths and a large proportion of stillbirths in the third trimester are related to this condition.\u003csup\u003e1-3\u003c/sup\u003e Reported rates of intrapartum-related neonatal encephalopathy vary, with estimates ranging from 1\u0026ndash;3 per 1,000 in high-income countries to significantly higher in low-resource settings.\u003csup\u003e3,4\u003c/sup\u003e Despite widespread use of cardiotocography (CTG) for fetal surveillance, outcomes have not improved, while cesarean rates and medicolegal costs have surged.\u003csup\u003e2,5,6\u003c/sup\u003e\u003c/p\u003e\n\u003cp\u003eMachine learning models for CTG interpretation have shown limited impact, often hindered by small datasets, inadequate adjustment for confounding, and exclusion of signal segments deemed \u0026lsquo;noisy\u0026rsquo; or \u0026lsquo;artifact\u0026rsquo;\u0026mdash;segments that may still contain clinically relevant patterns.\u003csup\u003e7-9\u003c/sup\u003e Excessive preprocessing may obscure associations with asphyxia risk. The rarity of asphyxia further limits statistical power, and a lack of causal discovery risks misattributing predictive importance to spurious features.\u003csup\u003e7,10,11\u003c/sup\u003e\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eFetal heart rate (FHR) is typically acquired by CTG monitors at ~4 Hz using either Doppler ultrasound or fetal scalp electrode (electrocardiography, ECG). The former captures mechanical motion while the latter directly records cardiac electrical activity, generally offering greater signal fidelity during active labor. Mode selection is subjective by clinician\u0026mdash;ECG use generally follows failed ultrasound acquisition during labor. Signal dropout is stored as zeros in backend systems and displayed as trace gaps. Although proprietary algorithms enhance visual readability through autocorrelation and filtering, this may conceal clinically important data loss.\u003csup\u003e7,12\u003c/sup\u003e Dropout differs by acquisition mode: ultrasound is more prone to signal loss due to maternal or fetal factors, while ECG\u0026mdash;despite directly capturing cardiac activity\u0026mdash;still exhibits dropout, a counterintuitive finding that may have a relationship with underlying fetal compromise. FDA reports have cited signal quality failures in adverse outcomes.\u003csup\u003e8\u003c/sup\u003e\u003c/p\u003e\n\u003cp\u003eAcross disciplines, the challenge of distinguishing signal loss due to technical limitations versus physiological compromise is shared in many biomedical domains\u0026mdash;from ECG and EEG to wearables and implantables.\u003csup\u003e13\u003c/sup\u003e Our findings contribute to this growing body of work, demonstrating that signal dropout can act as a latent biomarker, especially when stratified by acquisition context. This work also speaks to broader trends in AI-driven health monitoring: the transition from pattern recognition to causal, real-time inference from raw and imperfect signals.\u003csup\u003e11\u003c/sup\u003e\u003c/p\u003e\n\u003cp\u003eAs CTG becomes increasingly digitized, signal quality itself may hold prognostic value. We applied causal machine learning to assess whether mode-specific FHR dropout predicts perinatal asphyxia. By treating signal dropout as an informative exposure rather than dismissing it as nuisance noise, we aim to support the development of real-time, mode-aware decision-support tools in fetal monitoring. By uncovering underlying causal drivers using causal machine learning\u0026mdash;including those embedded in signal quality\u0026mdash;we seek to inform more robust and generalizable prediction models for real-world clinical deployment.\u003c/p\u003e"},{"header":"MATERIALS AND METHODS","content":"\u003ch4\u003eStudy Design and Data Sources\u003c/h4\u003e\n\u003cp\u003eThis multicenter cohort study included 52,757 laboring women at \u0026ge;36 weeks\u0026rsquo; gestation across three international datasets (Figure 1). The training cohort consisted of 36,792 births from Mercy Hospital for Women, Melbourne (2010\u0026ndash;2021).This is a tertiary hospoital.\u0026nbsp;Inclusion required \u0026ge;15 minutes of in-labor CTG data within the final 60 minutes before birth, regardless of labor stage or delivery mode, to ensure inclusion of both vaginal and cesarean births. This aligns with standards in computational fetal monitoring studies.\u003csup\u003e14,15\u003c/sup\u003e Multiple gestations and congenital anomalies were excluded. Active labor was confirmed by midwifery and obstetric annotations.\u003c/p\u003e\n\u003ch4\u003eData Extraction and Preprocessing\u003c/h4\u003e\n\u003cp\u003eCTG signals were extracted from Philips IntelliSpace\u0026trade; (Philips, Amsterdam) with technical support from Philips. Traces were linked to clinical data from the Birthing Outcomes System (BOS), hospital morbidity records, and paper charts as needed\u003csup\u003e7\u003c/sup\u003e. All data were stored securely using the University of Melbourne\u0026rsquo;s MediaFlux\u0026trade; platform, with analysis conducted on the Spartan high-performance computing cluster.\u003c/p\u003e\n\u003ch4\u003eExternal Validation Datasets\u003c/h4\u003e\n\u003cp\u003eValidation was performed on two cohorts. The first comprised 552 cases from a European dataset -the Czech CTU-UHB database (2010\u0026ndash;2012), an open-access European dataset with linked clinical metadata.\u003csup\u003e16\u003c/sup\u003e The second included 15,413 births from Werribee Mercy Hospital (2010\u0026ndash;2021),a community hospital in Australia. The data was processed identically to the training dataset. Both required \u0026ge;15 minutes of final-hour labor CTG data and excluded multiple gestations and congenital anomalies.\u003c/p\u003e\n\u003ch4\u003eQuality Control\u003c/h4\u003e\n\u003cp\u003eCTG recordings with \u0026gt;90% signal dropout were excluded due to their clinical unreliability.\u003c/p\u003e\n\u003ch4\u003eExposure Classification\u003c/h4\u003e\n\u003cp\u003eThe primary exposure was a high level of missing valid fetal heart rate (FHR) signal, specifically of the \u003cem\u003esignal dropout\u003c/em\u003e type\u0026mdash;defined as the cumulative proportion of zero-valued data points (missing valid biosignal)recorded during active signal acquisition in a live fetus. In Philips systems, these dropouts appear as visual gaps in the CTG trace; however, proprietary signal-processing algorithms (e.g., autocorrelation-based smoothing) may reduce their visibility. Dropout was calculated at a sampling frequency of 4 Hz over the final hour of labor. Based on Delphi consensus from a multidisciplinary team of clinicians, data scientists, and signal-processing experts, high dropout was predefined as \u0026gt;30% signal loss.\u003csup\u003e7\u003c/sup\u003e Maternal\u0026ndash;FHR coincidence was not assessed in this analysis.\u003c/p\u003e\n\u003cp\u003eParticipants were categorized into four exposure groups based on monitoring mode (ultrasound vs. ECG) and dropout level (low vs. high):\u003c/p\u003e\n\u003col\u003e\n \u003cli\u003eLow dropout in ultrasound mode (reference)\u003c/li\u003e\n \u003cli\u003eHigh dropout in ultrasound mode\u003c/li\u003e\n \u003cli\u003eLow dropout in ECG mode\u003c/li\u003e\n \u003cli\u003eHigh dropout in ECG mode\u003c/li\u003e\n\u003c/ol\u003e\n\u003cp\u003eMode classification was based on the presence and duration of ECG signal. Patients were assigned to the ECG group if \u0026ge;10 minutes of ECG was recorded, minimizing misclassification from brief mode switches. Those with \u0026lt;10 minutes were classified as ultrasound. This approach enabled mode-specific dropout analysis while reducing bias from clinician-initiated switching.\u003c/p\u003e\n\u003ch4\u003eOutcome Measures\u003c/h4\u003e\n\u003cp\u003eThe primary outcome was perinatal asphyxia, defined using a composite developed by multidisciplinary Delphi consensus and aligned with international standards. Criteria included one or more of: resuscitation at 10 minutes, therapeutic hypothermia, Apgar score \u0026le;6 at 10 minutes or \u0026le;4 at 5 minutes, cord pH \u0026lt;7.05, base excess \u0026lt;\u0026ndash;12 mmol/L, NICU admission, or clinically (including seizures) /imaging-confirmed hypoxic\u0026ndash;ischaemic encephalopathy (HIE).\u003c/p\u003e\n\u003ch4\u003eCausal Machine Learning Methods\u003c/h4\u003e\n\u003cp\u003eConfounder adjustment was guided by directed acyclic graphs (DAGs) (see Supplement). A doubly robust framework combining propensity score modeling and outcome regression was used to assess the causal effect of mode-specific dropout. All confounders used in causal models are listed in supplementary material and include maternal demographics, obstetric risk factors, labor characteristics, and mode of delivery. Propensity scores were estimated via multinomial logistic regression using maternal, fetal, and intrapartum variables. Inverse probability weights (IPWs), stabilized by clipping between 0.01 and 0.99, were used in weighted logistic regression. Covariate balance was assessed using standardized mean differences. HC3 robust standard errors accounted for heteroskedasticity. Sensitivity analyses were conducted on both complete case and imputed datasets.\u003c/p\u003e\n\u003cp\u003eTargeted Maximum Likelihood Estimation (TMLE) refined causal estimates for ECG High Dropout vs. Ultrasound Low Dropout. TMLE combined logistic regression and gradient boosting for exposure and outcome modeling. Ridge regularization was applied for logistic regression; the gradient boosting model used a learning rate of 0.1, max depth 3, and 100 estimators. Categorical variables were one-hot encoded; continuous variables (e.g., BMI) were log-transformed and standardized. Additional engineered features, such as labor duration ratios, were included. TMLE used a clever covariate (H) for targeted updates, and influence-curve\u0026ndash;based standard errors (via the \u003cem\u003ezepid\u003c/em\u003e package) for confidence intervals.\u0026nbsp;No additional fetal heart rate features (e.g., decelerations or variability) were modeled directly to isolate the effect of dropout.\u003c/p\u003e\n\u003ch4\u003eModel Validation and Sensitivity Analysis\u003c/h4\u003e\n\u003cp\u003eRobustness was evaluated across dropout thresholds using likelihood ratio tests (LRTs) and E-values for unmeasured confounding. Models were stratified by labor stage and tested on a 20% stratified holdout set from the training cohort. External validation was conducted using both the Czech and Werribee datasets.\u003c/p\u003e\n\u003ch4\u003eSurvival Analysis\u003c/h4\u003e\n\u003cp\u003eTime-to-event analysis assessed asphyxia risk from second-stage labor onset to event occurrence. \u0026quot;Survival\u0026quot; denoted remaining free from asphyxia. Exposure was categorized as a four-level composite of monitoring mode and dropout severity. Kaplan\u0026ndash;Meier curves were compared using log-rank tests. A weighted Cox proportional hazards model, adjusted using stabilized IPWs derived from a multinomial logistic regression, was used to account for confounding. Model discrimination was assessed using Harrell\u0026rsquo;s concordance index (C-index).\u0026nbsp;\u003c/p\u003e\n\u003cp\u003e\u003cem\u003eSample size calculation\u003c/em\u003e\u003c/p\u003e\n\u003cp\u003eConsistent with our prior study on CTG signal quality and perinatal outcomes\u0026mdash;conducted on a related but differently grouped cohort\u0026mdash;the current analysis builds on a pre-established dataset, including one additional large dataset, for which formal sample size justification had previously been undertaken\u003csup\u003e7\u003c/sup\u003e. Given the exploratory nature of this causal inference study and the rarity of perinatal asphyxia, we analyzed the full cohort (n = 52,757) to maximize statistical power for effect estimation, threshold modeling, and validation across independent populations.\u003c/p\u003e\n\u003ch4\u003eData Processing and Statistical Analysis\u003c/h4\u003e\n\u003cp\u003eAll analyses were performed using Python 3.6(including \u003cem\u003elifelines\u003c/em\u003e v0.30.0). and Stata 18.5. CTG data were extracted via Philips IntelliSpace\u0026trade; and managed on Spartan and MediaFlux\u0026trade; infrastructures. The detailed statistical analysis plan is enclosed in the supplementary material.\u003c/p\u003e\n\u003ch4\u003eHandling Missing Clinical Data\u003c/h4\u003e\n\u003cp\u003eMissing clinical variables were imputed using chained equations. Sensitivity analyses compared complete-case and imputed models. Maternal BMI had the highest missingness (6.32%, 95% CI: 6.06\u0026ndash;6.66%).\u003c/p\u003e\n\u003ch4\u003eEthical Approval and Reporting Standards\u003c/h4\u003e\n\u003cp\u003eThe study was approved by the institutional Human Research Ethics Committee (HREC #R2020-077). It adhered to a pre-specified statistical analysis plan and followed STROBE\u003csup\u003e17\u003c/sup\u003e and TRIPOD+AI guidelines(see checklists in Supplementary Materials).\u003csup\u003e18\u003c/sup\u003e\u0026nbsp;\u003c/p\u003e"},{"header":"RESULTS ","content":"\u003cp\u003eOf 36,792 laboring women screened in the training dataset, 32,242 met the inclusion criteria. There were 860 cases of perinatal asphyxia (2.67%, 95% CI: 2.49–2.85%) (Figure1). Table 1 shows baseline sample characteristics. Sample characteristics of the validation datasets are included in supplementary tables 1 and 2. Dropout patterns significantly varied by monitoring modality. Most cases were in the \u003cem\u003eUltrasound Low Dropout\u003c/em\u003e group (Group 0, 54.4%, 95% CI: 53.8%–54.9%), which had the lowest asphyxia rate (2.2%, 95% CI: 2.0%–2.4%). In contrast, the \u003cem\u003eECG High Dropout group\u003c/em\u003e (Group 3, 1.18%, 95% CI: 1.07%–1.31%) had the highest asphyxia rate (5.2%, 95% CI: 3.4%–7.9%)(p \u0026lt; 0.0001) (Figure 1).\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eMaternal and neonatal characteristics differed across groups.\u003cem\u003e\u0026nbsp;Maternal BMI\u003c/em\u003e was highest in ECG groups, with a median of 25.0 (IQR: 22.0–29.0) compared to 23.0 (IQR: 21.0–27.0) in the US Low Dropout group (p \u0026lt; 0.0001). \u003cem\u003eGestational age\u003c/em\u003e was similar across groups (p = 0.09), but \u003cem\u003ebaby weight\u003c/em\u003e was significantly lower in ECG High Dropout (3280g, IQR: 2970–3610) compared to US Low Dropout (3410g, IQR: 3110–3710, p \u0026lt; 0.0001). \u003cem\u003eParity\u0026nbsp;\u003c/em\u003edistributions also varied, with nulliparous women being most common in the ECG Low Dropout group (64.7%) versus 47.5% in US Low Dropout (p \u0026lt; 0.0001). Figure 2 shows representative modal devices and representative outputs.\u003c/p\u003e\n\u003cp\u003e\u003cem\u003e\u003cu\u003eCausal Inference Analysis\u003c/u\u003e\u003c/em\u003e\u003c/p\u003e\n\u003cp\u003eLogistic regression with inverse probability weighting (IPW) was used to estimate the association between signal dropout and perinatal asphyxia, using \u003cem\u003eUltrasound Low Dropout\u003c/em\u003e as the reference group (Figure 3a). The highest risk of asphyxia was observed in the \u003cem\u003eECG High Dropout\u0026nbsp;\u003c/em\u003egroup (OR = 3.14, 95% CI: 2.20–4.48, p \u0026lt; 0.001). The \u003cem\u003eECG Low Dropout\u0026nbsp;\u003c/em\u003egroup also showed significantly increased risk (OR = 1.51, 95% CI: 1.32–1.71, p \u0026lt; 0.001). The \u003cem\u003eUltrasound High Dropout\u003c/em\u003e group showed a non-significant risk of asphyxia (OR = 1.25, 95% CI: 0.99–1.58, p = 0.059) (table 2, Figure 3a).\u003c/p\u003e\n\u003cp\u003eE-value analysis assessed robustness to unmeasured confounding. The \u003cem\u003eECG High Dropout\u003c/em\u003e group showed the greatest robustness (E-value = 5.73; lower bound = 3.82). \u003cem\u003eECG Low Dropout\u003c/em\u003e showed moderate robustness (E-value = 2.39), while \u003cem\u003eUltrasound High Dropout\u003c/em\u003e, which was not statistically significant, showed limited robustness (E-value = 1.81). A likelihood ratio test (LRT) confirmed that the inclusion of the 'dropout by mode' classification significantly improved model fitcompared to a model without it (χ² (1) = 25.58, p \u0026lt; 0.0001) supporting the relevance of mode-specific dropout classification in explaining variation in asphyxia risk.\u003c/p\u003e\n\u003cp\u003e\u003cem\u003e\u003cu\u003eCausal Machine Learning Analysis\u003c/u\u003e\u003c/em\u003e\u003cbr\u003e\u0026nbsp;Targeted Maximum Likelihood Estimation (TMLE) refined causal estimates for the two key groups (US Low Dropout as reference and ECG High Dropout with highest asphyxia risk) by integrating machine learning–based propensity score modeling (Gradient Boosting Classifier) and outcome prediction (Gradient Boosting Regressor) to address high-dimensional, nonlinear relationships. TMLE confirmed increased asphyxia risk in ECG High Dropout (adjusted OR = 2.08, 95% CI: 1.21–3.58) relative to US Low Dropout (Table 2). In crude analyses, ECG High Dropout was associated with a 5.0% absolute risk increase—one additional case of asphyxia for every 20 laboring women monitored. After TMLE adjustment, the absolute risk increase was 2.27%, or one extra case per 44 labor episodes (95% CI: 28–83).\u003c/p\u003e\n\u003cp\u003eWe applied TMLE-derived individual risk estimates to assess fetal risk reclassification in the training set. Clinically, cases were considered low risk if labor progressed without operative delivery, and high risk if operative birth (caesarean, forceps, or vacuum) occurred due to suspected fetal distress. TMLE reclassified 9.29% of cases managed as low-risk into a high-risk category based on elevated model-predicted asphyxia risk. Reclassification was significantly more frequent among ECG-monitored cases, suggesting dropout-informed causal modeling may identify high-risk fetuses missed by conventional assessment. TMLE analyses across dropout thresholds showed progressively increasing odds of asphyxia.\u0026nbsp;Dropout \u0026lt;15% was not significantly associated with asphyxia, while increasing dropout thresholds from ≥20% showed a dose-response relationship with rising risk: OR = 1.56 (95% CI: 1.16–2.09, p \u0026lt; 0.001) at 20%, increasing to OR = 2.08 (95% CI: 1.21–3.58, p = 0.01) at 30% (Figure 3a). SHAP analysis in Supplementary Figure 3 illustrates features driving ECG High Dropout and asphyxia risk.\u003c/p\u003e\n\u003cp\u003e\u003cem\u003e\u003cu\u003eSurvival Analysis (Figure 3b)\u003c/u\u003e\u003c/em\u003e\u003cbr\u003e\u0026nbsp;Kaplan–Meier estimates stratified by monitoring modality and signal dropout showed the highest asphyxia-free survival in the Ultrasound Low Dropout group, and the steepest decline in the ECG High Dropout group. Differences were statistically significant (log-rank χ² = 43.6, p \u0026lt; 0.0001).\u003cbr\u003e\u0026nbsp;To adjust for confounding, we applied an inverse probability–weighted Cox proportional hazards model using stabilized weights from a multinomial treatment model. In the fully adjusted model, the hazard of asphyxia was significantly higher in ECG High Dropout versus US Low Dropout (HR = 4.76, 95% CI: 2.98–7.62, p \u0026lt; 0.0001). Elevated hazards were also observed for Ultrasound High Dropout (HR = 1.90, 95% CI: 1.39–2.60, p \u0026lt; 0.0001) and ECG Low Dropout (HR = 1.42, 95% CI: 1.20–1.69, p \u0026lt; 0.0001) (Supplementary Table 3). Harrell’s C-index was 0.71, indicating moderate predictive accuracy. Figure 3b displays Kaplan–Meier and IPW-adjusted Cox curves by monitoring modality and dropout level.\u003c/p\u003e\n\u003cp\u003e\u003cem\u003e\u003cu\u003eExternal Validation\u003c/u\u003e\u003c/em\u003e\u003c/p\u003e\n\u003cp\u003eExternal validation using the Czech CTU-UHB and Werribee datasets demonstrated similar trends to the model generation cohort(table 3)\u003c/p\u003e\n\u003cp\u003e\u003cem\u003eValidation on Czech dataset\u003c/em\u003e\u003c/p\u003e\n\u003cp\u003eCausal inference analysis using inverse probability weighting (IPW) demonstrated that the \u003cem\u003eECG High Dropout\u0026nbsp;\u003c/em\u003egroup was significantly associated with an increased risk of perinatal asphyxia (OR = 4.02, 95% CI: 1.41–11.44, p = 0.009) compared to the \u003cem\u003eUltrasound Low Dropout\u003c/em\u003e reference group. The ECG \u003cem\u003eLow Dropout\u003c/em\u003e group showed a paradoxical reduction in asphyxia risk (OR = 0.24, 95% CI: 0.06–0.96, p = 0.043), which may reflect differing case mix, small sample size, or differences in mode-switching thresholds between centers. The \u003cem\u003eUltrasound High Dropout\u003c/em\u003e group showed a non-significant trend toward increased risk (OR = 1.90, 95% CI: 0.94–3.82, p = 0.073). Causal machine learning analysis using targeted maximum likelihood estimation (TMLE) produced consistent findings, estimating an asphyxia odds ratio of 2.83 (95% CI: 2.13–16.36) for the \u003cem\u003eECG High Dropout\u003c/em\u003e group compared to the reference \u003cem\u003eUltrasound Low Dropout\u003c/em\u003e.\u003c/p\u003e\n\u003cp\u003e\u003cem\u003eValidation on Werribee dataset\u0026nbsp;\u003c/em\u003e\u003c/p\u003e\n\u003cp\u003eCausal inference analysis using inverse probability weighting (IPW) regression model showed that the \u003cem\u003eECG High Dropout\u0026nbsp;\u003c/em\u003egroup was significantly associated with increased asphyxia risk (OR = 2.22, 95% CI: 1.61–3.05, p \u0026lt; 0.001), followed by the \u003cem\u003eECG Low Dropout\u0026nbsp;\u003c/em\u003egroup (OR = 1.57, 95% CI: 1.31–1.89, p \u0026lt; 0.001). The \u003cem\u003eUltrasound High Dropout\u0026nbsp;\u003c/em\u003egroup was not significantly associated with asphyxia (OR = 0.69, 95% CI: 0.33–1.46, p = 0.33), using \u003cem\u003eUltrasound Low Dropout\u003c/em\u003e as the reference. Causal machine learning using targeted maximum likelihood estimation (TMLE) showed that the \u003cem\u003eECG High Dropout\u003c/em\u003e group showed a significant increase in asphyxia risk (OR = 1.91, 95% CI: 1.11–3.28) compared to the reference \u003cem\u003eUltrasound Low Dropout\u003c/em\u003e.\u003c/p\u003e"},{"header":"DISCUSSION","content":"\u003cp\u003e\u003cem\u003e\u003cu\u003ePrincipal Findings \u0026nbsp;\u003c/u\u003e\u003c/em\u003e\u003c/p\u003e\n\u003cp\u003eIn this multicohort study of over 50,000 births, we show that mode-specific FHR signal dropout—typically overlooked in CTG interpretation—is strongly and causally associated with perinatal asphyxia, particularly in ECG-based monitoring. ECG High Dropout carried the highest risk (IPW OR = 3.14; 95% CI: 2.20–4.48), followed by ECG Low Dropout (OR = 1.51; 95% CI: 1.32–1.71). Ultrasound High Dropout was not significant, reinforcing the hypothesis that dropout in ECG—typically used in advanced labor—may more directly reflect underlying physiological compromise rather than technical factors. TMLE confirmed a dose-response relationship, with dropout ≥30% yielding an OR of 2.08 and an absolute risk increase of one asphyxia case per 44 women. TMLE-based risk profiling reclassified nearly 10% of clinically low-risk cases to high risk. Findings were consistent across external cohorts (Czech: IPW OR = 4.02, TMLE OR = 2.83; Werribee: IPW OR = 2.22, TMLE OR = 1.91), reinforcing the prognostic value of ECG dropout.\u003c/p\u003e\n\u003cp\u003e\u003cem\u003e\u003cu\u003eResults in the Context of What is Known\u003c/u\u003e\u003c/em\u003e\u003c/p\u003e\n\u003cp\u003eSignal dropout in ECG-mode CTG is a strong, clinically meaningful predictor of asphyxia—challenging the assumption that it reflects only technical or user error. Mode selection is subjective, and dropout may get hidden due to interface design. Our findings support making dropout metrics visible. Rather than default mode switching, decisions should be guided by dropout severity: US-mode dropout \u0026gt;30% may warrant switching to ECG, but persistent high dropout in ECG-mode indicates tripled asphyxia risk and should prompt closer surveillance\u003c/p\u003e\n\u003cp\u003e\u003cem\u003e\u003cu\u003eClinical Implications\u003c/u\u003e\u003c/em\u003e\u003c/p\u003e\n\u003cp\u003eA major strength of this study is the use of causal methods—IPW and TMLE—to estimate the effect of mode-specific dropout. Dropout is non-random and influenced by clinical factors; failure to adjust for this can bias observational analyses. Our approach improved covariate balance and model fit (χ²(1) = 16.2, p = 0.00006), with E-values (ECG High Dropout: 5.89) supporting robustness to unmeasured confounding.\u003c/p\u003e\n\u003cp\u003eUnlike traditional machine learning, which may overfit and overlook causal pathways, causal ML methods like TMLE explicitly model exposure–outcome relationships to enhance generalizability.\u003csup\u003e11,19-21\u003c/sup\u003e We used DAGs to select confounders and validated findings across two independent cohorts. Prior models, such as McCoy et al., excluded high-artifact CTGs (\u0026gt;30% dropout), discarding over one-third of cases.\u003csup\u003e22\u003c/sup\u003e While their deep learning model achieved high AUROC (0.85 for pH \u0026lt;7.05), it ignored dropout as a signal. Moreover, pH \u0026lt;7 alone does not reliably stratify neonatal outcomes, as most infants with pH between 6.9 and 7 survive without neurodevelopmental impairment.\u003csup\u003e23\u003c/sup\u003e In contrast, we treat signal dropout as a meaningful exposure with prognostic value and applied a more holistic, clinically accepted composite outcome aligned with established definitions of neurodevelopmental risk.\u003csup\u003e5,24,25\u003c/sup\u003e While our model focused on signal dropout, it did not incorporate contemporaneous CTG pattern interpretation (e.g., decelerations), which may offer complementary prognostic information in future models.\u003c/p\u003e\n\u003cp\u003e\u003cem\u003e\u003cu\u003eResearch Implications\u0026nbsp;\u003c/u\u003e\u003c/em\u003e\u003c/p\u003e\n\u003cp\u003eBy integrating mode-specific dropout into causal ML, we demonstrate its utility in identifying high-risk cases missed by conventional approaches. This supports real-time monitoring tools that adapt to dropout severity, aligning with current recommendations for actionable, causally grounded AI in healthcare.\u003csup\u003e11,19-21\u003c/sup\u003e Beyond obstetrics, our findings resonate with challenges in other domains that rely on physiological time series—such as ECG, EEG, and wearable biosensors—where distinguishing noise from latent risk is critical for advancing real-time, interpretable AI applications.\u003c/p\u003e\n\u003cp\u003e\u003cem\u003e\u003cu\u003e\u0026nbsp;Strengths and Limitations\u0026nbsp;\u003c/u\u003e\u003c/em\u003e\u003c/p\u003e\n\u003cp\u003eOur study leverages one of the largest digitized CTG datasets, with validation across multiple cohorts, enhancing generalizability. The integration of causal inference and machine learning improves robustness.\u003csup\u003e12,19,20,26,27\u003c/sup\u003e A limitation is the lack of data on real-time clinician-driven mode-switching decisions. However, consistent validation trends across Czech CTU-UHB and Werribee datasets reinforce the observed associations. Implementation in other settings may face logistical challenges, but prospective validation is essential to assess whether integrating dropout monitoring improves neonatal outcomes.\u0026nbsp;We acknowledge the lack of data on real-time clinician-driven mode-switching or intrauterine resuscitation interventions, which may confound the relationship between dropout and outcome\u003c/p\u003e\n\u003cp\u003eWhile our framework enhances risk estimation, clinical deployment would require real-time implementation within a comprehensive decision-support model. Future extensions may adapt approaches such as the AI-ECG sex discordance score, which identifies subtle, clinician-invisible signal shifts—analogous to dropout thresholds in our study.\u003csup\u003e28\u003c/sup\u003eHowever, developers must also address known demographic disparities in signal-based AI, as shown in recent ECG heart failure models, where performance biases were especially pronounced among young Black women.\u003csup\u003e29\u003c/sup\u003e This study also contributes to emerging efforts across disciplines to better utilize ‘imperfect’ or artifact-prone physiological data streams—turning what was once discarded into features that enhance clinical relevance and model transparency.\u003c/p\u003e"},{"header":"Conclusion","content":"\u003cp\u003eMode-specific FHR dropout\u0026mdash;particularly in ECG-mode CTG\u0026mdash;is a clinically significant predictor of perinatal asphyxia. Rather than being excluded or ignored, signal quality should be incorporated into predictive modeling. Integrating signal metrics into ensemble models may enhance fetal surveillance and should be prioritized in future validation efforts.\u003c/p\u003e"},{"header":"Declarations","content":"\u003cp\u003e\u003cstrong\u003eDisclosure/Conflict of interest:\u003c/strong\u003e\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eNone of the authors have a direct conflict of interest. FB and EK are directors and shareholders of Kali Healthcare, a company that is commercialising a wearable fetal monitoring device. \u0026nbsp;MP is a shareholder of Kali Healthcare. Rest of the authors have nothing to disclose.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eFunding\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eFB is supported by a Dame Kate Campbell Fellowship, a University of Melbourne Fellowship, and an National Health and Medical Research Council (NHMRC) Ideas Grant. DK receives postgraduate scholarship support from the NHMRC.Additional funding was provided by the Norman Beischer Medical Research Foundation. LM receives University of Melbourne Graeme Clark Institute and Melbourne Research scholarships. The authors were not precluded from accessing data in the study, and they accept responsibility to submit for publication.\u003c/strong\u003e\u003c/p\u003e\u003cp\u003e\u003cstrong\u003eAcknowledgments:\u003c/strong\u003e\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eThis research was supported by The University of Melbourne’s Research Computing Services\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eand the Petascale Campus Initiative.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eFunding:\u003c/strong\u003e\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eFB is supported by a Dame Kate Campbell Fellowship, a University of Melbourne Fellowship,\u003c/p\u003e\n\u003cp\u003eand an National Health and Medical Research Council (NHMRC) Ideas Grant. DK receives\u003c/p\u003e\n\u003cp\u003epostgraduate scholarship support from the NHMRC.Additional funding was provided by the\u003c/p\u003e\n\u003cp\u003eNorman Beischer Medical Research Foundation. LM receives University of Melbourne Graeme\u003c/p\u003e\n\u003cp\u003eClark Institute and Melbourne Research scholarships. The authors were not precluded from\u003c/p\u003e\n\u003cp\u003eaccessing data in the study, and they accept responsibility to submit for publication.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eAuthor contributions:\u003c/strong\u003e\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eDK: Conceptualization, Methodology, Investigation, Data curation, Writing – review \u0026amp; editing\u003c/p\u003e\n\u003cp\u003eVisualization, Funding acquisition, Writing – original draft\u003c/p\u003e\n\u003cp\u003eFB: Conceptualization, Methodology, Investigation, Funding acquisition, Project administration, Supervision, Writing – original draft\u003c/p\u003e\n\u003cp\u003eEM: Methodology, Visualization, Project administration, Supervision, Writing – review \u0026amp; editing\u003c/p\u003e\n\u003cp\u003eLM: Investigation, Data curation, Writing – review \u0026amp; editing\u003c/p\u003e\n\u003cp\u003eWriting – review \u0026amp; editing\u003c/p\u003e\n\u003cp\u003eEK: Supervision, Data curation, Writing – review \u0026amp; editing\u003c/p\u003e\n\u003cp\u003eWriting – review \u0026amp; editing\u003c/p\u003e\n\u003cp\u003eMP: Supervision, Writing – review \u0026amp; editing\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eCompeting interests:\u003c/strong\u003e\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eNone of the authors have a direct conflict of interest. FB and EK are directors and shareholders of Kali Healthcare, a company that is commercialising a wearable fetal monitoring device. \u0026nbsp;MP is a shareholder of Kali Healthcare. Rest of the authors have nothing to disclose.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eData and materials availability:\u003c/strong\u003e\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eAll individual-level data of the included cohort and modeling codes can be shared upon reasonable request to the corresponding author and completion of data transfer agreement forms.\u0026nbsp;\u003c/p\u003e"},{"header":"References","content":"\u003col\u003e\n\u003cli\u003eMota-Rojas, D.\u003cem\u003e, et al.\u003c/em\u003e Pathophysiology of perinatal asphyxia in humans and animal models. \u003cem\u003eBiomedicines\u003c/em\u003e \u003cstrong\u003e10\u003c/strong\u003e, 347 (2022).\u003c/li\u003e\n\u003cli\u003eResolution, N.H.S. Annual Report and Accounts 2019/20. (NHS Resolution, UK, 2020).\u003c/li\u003e\n\u003cli\u003eSchwartz, N. \u0026amp; Young, B.K. Intrapartum fetal monitoring today. (2006).\u003c/li\u003e\n\u003cli\u003eSchiermeier, S.\u003cem\u003e, et al.\u003c/em\u003e Sensitivity and specificity of intrapartum computerised FIGO criteria for cardiotocography and fetal scalp pH during labour: multicentre, observational study. \u003cem\u003eBjog\u003c/em\u003e \u003cstrong\u003e115\u003c/strong\u003e, 1557-1563 (2008).\u003c/li\u003e\n\u003cli\u003eBrocklehurst, P.\u003cem\u003e, et al.\u003c/em\u003e Computerised interpretation of fetal heart rate during labour (INFANT): a randomised controlled trial. \u003cem\u003eThe Lancet\u003c/em\u003e \u003cstrong\u003e389\u003c/strong\u003e, 1719-1729 (2017).\u003c/li\u003e\n\u003cli\u003eAyres-de-Campos, D., Spong, C.Y. \u0026amp; Chandraharan, E. FIGO consensus guidelines on intrapartum fetal monitoring: Cardiotocography. \u003cem\u003eInternational Journal of Gynecology \u0026amp; Obstetrics\u003c/em\u003e \u003cstrong\u003e131\u003c/strong\u003e, 13-24 (2015).\u003c/li\u003e\n\u003cli\u003eKarmakar, D.\u003cem\u003e, et al.\u003c/em\u003e Impact of missing electronic fetal monitoring signals on perinatal asphyxia: a multicohort analysis. \u003cem\u003eNPJ Digit Med\u003c/em\u003e \u003cstrong\u003e8\u003c/strong\u003e, 233 (2025).\u003c/li\u003e\n\u003cli\u003eKiely, D.J., Oppenheimer, L.W. \u0026amp; Dornan, J.C. Unrecognized maternal heart rate artefact in cases of perinatal mortality reported to the United States Food and Drug Administration from 2009 to 2019: a critical patient safety issue. \u003cem\u003eBMC Pregnancy and Childbirth\u003c/em\u003e \u003cstrong\u003e19\u003c/strong\u003e, 1-10 (2019).\u003c/li\u003e\n\u003cli\u003eSignorini, M.G. \u0026amp; Magenes, G. Advanced signal processing techniques for CTG analysis. in \u003cem\u003eXIV Mediterranean Conference on Medical and Biological Engineering and Computing 2016: MEDICON 2016, March 31st-April 2nd 2016, Paphos, Cyprus\u003c/em\u003e 1205-1210 (Springer, 2016).\u003c/li\u003e\n\u003cli\u003eO\u0026rsquo;Sullivan, M.E.\u003cem\u003e, et al.\u003c/em\u003e Challenges of developing robust AI for intrapartum fetal heart rate monitoring. \u003cem\u003eFrontiers in Artificial Intelligence\u003c/em\u003e \u003cstrong\u003e4\u003c/strong\u003e, 765210 (2021).\u003c/li\u003e\n\u003cli\u003eProsperi, M.\u003cem\u003e, et al.\u003c/em\u003e Causal inference and counterfactual prediction in machine learning for actionable healthcare. \u003cem\u003eNature Machine Intelligence\u003c/em\u003e \u003cstrong\u003e2\u003c/strong\u003e, 369-375 (2020).\u003c/li\u003e\n\u003cli\u003eBlakely, T., Lynch, J., Simons, K., Bentley, R. \u0026amp; Rose, S. Reflection on modern methods: when worlds collide\u0026mdash;prediction, machine learning and causal inference. \u003cem\u003eInternational Journal of Epidemiology\u003c/em\u003e \u003cstrong\u003e49\u003c/strong\u003e, 2058-2064 (2019).\u003c/li\u003e\n\u003cli\u003eClifford, G.D., Behar, J., Li, Q. \u0026amp; Rezek, I. Signal quality indices and data fusion for determining clinical acceptability of electrocardiograms. \u003cem\u003ePhysiol Meas\u003c/em\u003e \u003cstrong\u003e33\u003c/strong\u003e, 1419-1433 (2012).\u003c/li\u003e\n\u003cli\u003eMendis, L., Palaniswami, M., Brownfoot, F. \u0026amp; Keenan, E. Computerised cardiotocography analysis for the automated detection of fetal compromise during labour: a review. \u003cem\u003eBioengineering\u003c/em\u003e \u003cstrong\u003e10\u003c/strong\u003e, 1007 (2023).\u003c/li\u003e\n\u003cli\u003ePetrozziello, A., Redman, C.W., Papageorghiou, A.T., Jordanov, I. \u0026amp; Georgieva, A. Multimodal convolutional neural networks to detect fetal compromise during labor and delivery. \u003cem\u003eIEEE Access\u003c/em\u003e \u003cstrong\u003e7\u003c/strong\u003e, 112026-112036 (2019).\u003c/li\u003e\n\u003cli\u003eChud\u0026aacute;ček, V.\u003cem\u003e, et al.\u003c/em\u003e Open access intrapartum CTG database. \u003cem\u003eBMC pregnancy and childbirth\u003c/em\u003e \u003cstrong\u003e14\u003c/strong\u003e, 1-12 (2014).\u003c/li\u003e\n\u003cli\u003eVon Elm, E.\u003cem\u003e, et al.\u003c/em\u003e The Strengthening the Reporting of Observational Studies in Epidemiology (STROBE) statement: guidelines for reporting observational studies. \u003cem\u003eThe lancet\u003c/em\u003e \u003cstrong\u003e370\u003c/strong\u003e, 1453-1457 (2007).\u003c/li\u003e\n\u003cli\u003eCollins, G.S.\u003cem\u003e, et al.\u003c/em\u003e TRIPOD+ AI statement: updated guidance for reporting clinical prediction models that use regression or machine learning methods. \u003cem\u003ebmj\u003c/em\u003e \u003cstrong\u003e385\u003c/strong\u003e(2024).\u003c/li\u003e\n\u003cli\u003eHern\u0026aacute;n, M.A. \u0026amp; Robins, J.M. Using Big Data to Emulate a Target Trial When a Randomized Trial Is Not Available. \u003cem\u003eAm J Epidemiol\u003c/em\u003e \u003cstrong\u003e183\u003c/strong\u003e, 758-764 (2016).\u003c/li\u003e\n\u003cli\u003eHern\u0026aacute;n, M.A., Wang, W. \u0026amp; Leaf, D.E. Target trial emulation: a framework for causal inference from observational data. \u003cem\u003eJama\u003c/em\u003e \u003cstrong\u003e328\u003c/strong\u003e, 2446-2447 (2022).\u003c/li\u003e\n\u003cli\u003eNevo, D., Ogino, S. \u0026amp; Wang, M. Reflection on modern methods: causal inference considerations for heterogeneous disease etiology. \u003cem\u003eInternational Journal of Epidemiology\u003c/em\u003e \u003cstrong\u003e50\u003c/strong\u003e, 1030-1037 (2021).\u003c/li\u003e\n\u003cli\u003eMcCoy, J.A.\u003cem\u003e, et al.\u003c/em\u003e Intrapartum electronic fetal heart rate monitoring to predict acidemia at birth with the use of deep learning. \u003cem\u003eAmerican Journal of Obstetrics and Gynecology\u003c/em\u003e (2024).\u003c/li\u003e\n\u003cli\u003eLau, S.L.\u003cem\u003e, et al.\u003c/em\u003e Neonatal outcome of infants with umbilical cord arterial pH less than 7. \u003cem\u003eActa Obstet Gynecol Scand\u003c/em\u003e \u003cstrong\u003e102\u003c/strong\u003e, 174-180 (2023).\u003c/li\u003e\n\u003cli\u003eLee, A.C.\u003cem\u003e, et al.\u003c/em\u003e Intrapartum-related neonatal encephalopathy incidence and impairment at regional and global levels for 2010 with trends from 1990. \u003cem\u003ePediatric research\u003c/em\u003e \u003cstrong\u003e74\u003c/strong\u003e, 50-72 (2013).\u003c/li\u003e\n\u003cli\u003eRazak, A.\u003cem\u003e, et al.\u003c/em\u003e Early Neurodevelopmental Assessments for Predicting Long-Term Outcomes in Infants at High Risk of Cerebral Palsy. \u003cem\u003eJAMA Network Open\u003c/em\u003e \u003cstrong\u003e7\u003c/strong\u003e, e2413550-e2413550 (2024).\u003c/li\u003e\n\u003cli\u003eLuque‐Fernandez, M.A., Schomaker, M., Rachet, B. \u0026amp; Schnitzer, M.E. Targeted maximum likelihood estimation for a binary treatment: A tutorial. \u003cem\u003eStatistics in medicine\u003c/em\u003e \u003cstrong\u003e37\u003c/strong\u003e, 2530-2546 (2018).\u003c/li\u003e\n\u003cli\u003eSmith, M.J., Phillips, R.V., Luque-Fernandez, M.A. \u0026amp; Maringe, C. Application of targeted maximum likelihood estimation in public health and epidemiological studies: a systematic review. \u003cem\u003eAnnals of epidemiology\u003c/em\u003e \u003cstrong\u003e86\u003c/strong\u003e, 34-48. e28 (2023).\u003c/li\u003e\n\u003cli\u003eSau, A.\u003cem\u003e, et al.\u003c/em\u003e Artificial intelligence-enhanced electrocardiography for the identification of a sex-related cardiovascular risk continuum: a retrospective cohort study. \u003cem\u003eThe Lancet Digital Health\u003c/em\u003e \u003cstrong\u003e7\u003c/strong\u003e, e184-e194 (2025).\u003c/li\u003e\n\u003cli\u003eKaur, D.\u003cem\u003e, et al.\u003c/em\u003e Race, sex, and age disparities in the performance of ECG deep learning models predicting heart failure. \u003cem\u003eCirculation: Heart Failure\u003c/em\u003e \u003cstrong\u003e17\u003c/strong\u003e, e010879 (2024).\u003c/li\u003e\n\u003c/ol\u003e"},{"header":"Tables","content":"\u003cp\u003e\u003cstrong\u003eTable 1: Baseline characteristics by mode specific dropout levels\u003c/strong\u003e\u003c/p\u003e\n\u003ctable border=\"0\" cellspacing=\"0\" cellpadding=\"0\" width=\"662\"\u003e\n \u003ctbody\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 161px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 104px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eUS mode, low dropout\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 94px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eUS mode, high dropout\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 104px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eECG mode, low dropout\u0026nbsp;\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 94px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eECG mode, high dropout\u0026nbsp;\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 104px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eP-value\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 161px;\"\u003e\n \u003cp\u003eNumber of records, % proportion, (95%CI)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 104px;\"\u003e\n \u003cp\u003e17533, (54.38,53.83-54.92)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 94px;\"\u003e\n \u003cp\u003e2363, (7.33,7.05-7.62)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 104px;\"\u003e\n \u003cp\u003e11964, (37.11,36.58-37.64)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 94px;\"\u003e\n \u003cp\u003e382(1.18,1,07-1.31)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 104px;\"\u003e\n \u003cp\u003e0.00\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 161px;\"\u003e\n \u003cp\u003eProportion of CTG signal by US mode\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 104px;\"\u003e\n \u003cp\u003e1.00 (1.00-1.00)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 94px;\"\u003e\n \u003cp\u003e1.00 (1.00-1.00)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 104px;\"\u003e\n \u003cp\u003e0.04 (0.00-0.35)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 94px;\"\u003e\n \u003cp\u003e0.70 (0.57-0.83)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 104px;\"\u003e\n \u003cp\u003e0.00\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 161px;\"\u003e\n \u003cp\u003eMaternal age\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 104px;\"\u003e\n \u003cp\u003e32.00 (29.00-35.00)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 94px;\"\u003e\n \u003cp\u003e32.00 (29.00-35.00)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 104px;\"\u003e\n \u003cp\u003e32.00 (29.00-35.00)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 94px;\"\u003e\n \u003cp\u003e32.50 (29.00-36.00)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 104px;\"\u003e\n \u003cp\u003e0.00\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 161px;\"\u003e\n \u003cp\u003eGestational age\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 104px;\"\u003e\n \u003cp\u003e39.40 (38.50-40.30)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 94px;\"\u003e\n \u003cp\u003e39.40 (38.50-40.30)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 104px;\"\u003e\n \u003cp\u003e39.50 (38.50-40.40)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 94px;\"\u003e\n \u003cp\u003e39.40 (38.40-40.30)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 104px;\"\u003e\n \u003cp\u003e0.09\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 161px;\"\u003e\n \u003cp\u003eMaternal BMI\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 104px;\"\u003e\n \u003cp\u003e23.00 (21.00-27.00)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 94px;\"\u003e\n \u003cp\u003e24.00 (21.00-28.00)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 104px;\"\u003e\n \u003cp\u003e25.00 (22.00-29.00)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 94px;\"\u003e\n \u003cp\u003e25.00 (22.00-29.00)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 104px;\"\u003e\n \u003cp\u003e\u0026lt;0.0001\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 161px;\"\u003e\n \u003cp\u003eBaby Weight(grams)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 104px;\"\u003e\n \u003cp\u003e3410.00 (3110.00-3710.00)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 94px;\"\u003e\n \u003cp\u003e3380.00 (3070.00-3690.00)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 104px;\"\u003e\n \u003cp\u003e3350.00 (3040.00-3660.00)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 94px;\"\u003e\n \u003cp\u003e3280.00 (2970.00-3610.00)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 104px;\"\u003e\n \u003cp\u003e\u0026lt;0.0001\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 161px;\"\u003e\n \u003cp\u003eParity\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 104px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 94px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 104px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 94px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 104px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 161px;\"\u003e\n \u003cp\u003e\u003cem\u003eNulliparous\u003c/em\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 104px;\"\u003e\n \u003cp\u003e47.5% (0.468 - 0.483)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 94px;\"\u003e\n \u003cp\u003e35.3% (0.334 - 0.373)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 104px;\"\u003e\n \u003cp\u003e64.7% (0.638 - 0.655)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 94px;\"\u003e\n \u003cp\u003e51.0% (0.460 - 0.560)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd rowspan=\"4\" valign=\"top\" style=\"width: 104px;\"\u003e\n \u003cp\u003e\u0026lt;0.0001\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 161px;\"\u003e\n \u003cp\u003e\u003cem\u003ePara 1\u003c/em\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 104px;\"\u003e\n \u003cp\u003e33.8% (0.331 - 0.345)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 94px;\"\u003e\n \u003cp\u003e46.0% (0.440 - 0.481)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 104px;\"\u003e\n \u003cp\u003e24.3% (0.236 - 0.251)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 94px;\"\u003e\n \u003cp\u003e35.6% (0.310 - 0.405)\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 161px;\"\u003e\n \u003cp\u003e\u003cem\u003ePara 2\u003c/em\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 104px;\"\u003e\n \u003cp\u003e12.9% (0.124 - 0.134)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 94px;\"\u003e\n \u003cp\u003e13.2% (0.119 - 0.147)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 104px;\"\u003e\n \u003cp\u003e7.7% (0.073 - 0.082)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 94px;\"\u003e\n \u003cp\u003e9.4% (0.069 - 0.128)\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 161px;\"\u003e\n \u003cp\u003e\u003cem\u003ePara 3 or higher\u003c/em\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 104px;\"\u003e\n \u003cp\u003e5.7% (0.054 - 0.061)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 94px;\"\u003e\n \u003cp\u003e5.4% (0.045 - 0.064)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 104px;\"\u003e\n \u003cp\u003e3.2% (0.029 - 0.036)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 94px;\"\u003e\n \u003cp\u003e3.9% (0.024 - 0.064)\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 161px;\"\u003e\n \u003cp\u003eBaby gender\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 104px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 94px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 104px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 94px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 104px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 161px;\"\u003e\n \u003cp\u003e\u003cem\u003eFemale/indeterminate\u003c/em\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 104px;\"\u003e\n \u003cp\u003e49.6% (0.489 - 0.503)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 94px;\"\u003e\n \u003cp\u003e50.0% (0.480 - 0.520)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 104px;\"\u003e\n \u003cp\u003e47.6% (0.467 - 0.484)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 94px;\"\u003e\n \u003cp\u003e47.9% (0.429 - 0.529)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd rowspan=\"2\" valign=\"top\" style=\"width: 104px;\"\u003e\n \u003cp\u003e0.00\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 161px;\"\u003e\n \u003cp\u003e\u003cem\u003eMale\u003c/em\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 104px;\"\u003e\n \u003cp\u003e50.4% (0.497 - 0.511)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 94px;\"\u003e\n \u003cp\u003e50.0% (0.480 - 0.520)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 104px;\"\u003e\n \u003cp\u003e52.4% (0.516 - 0.533)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 94px;\"\u003e\n \u003cp\u003e52.1% (0.471 - 0.571)\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 161px;\"\u003e\n \u003cp\u003ePrimary Obstetrics Diagnosis\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 104px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 94px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 104px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 94px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 104px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 161px;\"\u003e\n \u003cp\u003e\u003cem\u003eLow risk pregnancy\u003c/em\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 104px;\"\u003e\n \u003cp\u003e44.0% (0.433 - 0.448)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 94px;\"\u003e\n \u003cp\u003e60.0% (0.580 - 0.620)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 104px;\"\u003e\n \u003cp\u003e35.3% (0.344 - 0.361)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 94px;\"\u003e\n \u003cp\u003e48.4% (0.435 - 0.534)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd rowspan=\"5\" valign=\"top\" style=\"width: 104px;\"\u003e\n \u003cp\u003e\u0026lt;0.0001\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 161px;\"\u003e\n \u003cp\u003e\u003cem\u003eFetal compromise\u003c/em\u003e\u0026dagger;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 104px;\"\u003e\n \u003cp\u003e19.9% (0.193 - 0.205)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 94px;\"\u003e\n \u003cp\u003e12.0% (0.107 - 0.133)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 104px;\"\u003e\n \u003cp\u003e25.3% (0.246 - 0.261)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 94px;\"\u003e\n \u003cp\u003e16.0% (0.126 - 0.200)\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 161px;\"\u003e\n \u003cp\u003e\u003cem\u003eMiscellaneous fetal conditions\u003c/em\u003e\u0026dagger;\u0026dagger;\u003cem\u003e\u0026nbsp;\u003c/em\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 104px;\"\u003e\n \u003cp\u003e3.2% (0.029 - 0.034)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 94px;\"\u003e\n \u003cp\u003e1.7% (0.013 - 0.023)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 104px;\"\u003e\n \u003cp\u003e3.7% (0.033 - 0.040)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 94px;\"\u003e\n \u003cp\u003e1.8% (0.009 - 0.037)\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 161px;\"\u003e\n \u003cp\u003e\u003cem\u003eMaternal medical conditions\u003c/em\u003e\u003cem\u003e\u0026sect;\u003c/em\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 104px;\"\u003e\n \u003cp\u003e16.9% (0.164 - 0.175)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 94px;\"\u003e\n \u003cp\u003e14.3% (0.129 - 0.157)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 104px;\"\u003e\n \u003cp\u003e19.2% (0.185 - 0.200)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 94px;\"\u003e\n \u003cp\u003e19.1% (0.155 - 0.234)\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 161px;\"\u003e\n \u003cp\u003e\u003cem\u003ePre labor rupture of membranes\u003c/em\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 104px;\"\u003e\n \u003cp\u003e12.1% (0.117 - 0.126)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 94px;\"\u003e\n \u003cp\u003e9.2% (0.081 - 0.104)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 104px;\"\u003e\n \u003cp\u003e13.0% (0.124 - 0.136)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 94px;\"\u003e\n \u003cp\u003e11.5% (0.087 - 0.151)\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 161px;\"\u003e\n \u003cp\u003eOnset of labor\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 104px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 94px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 104px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 94px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 104px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 161px;\"\u003e\n \u003cp\u003e\u003cem\u003eInduced\u003c/em\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 104px;\"\u003e\n \u003cp\u003e49.2% (0.485 - 0.500)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 94px;\"\u003e\n \u003cp\u003e29.2% (0.274 - 0.311)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 104px;\"\u003e\n \u003cp\u003e60.8% (0.600 - 0.617)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 94px;\"\u003e\n \u003cp\u003e43.5% (0.386 - 0.485)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd rowspan=\"2\" valign=\"top\" style=\"width: 104px;\"\u003e\n \u003cp\u003e\u0026lt;0.0001\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 161px;\"\u003e\n \u003cp\u003e\u003cem\u003eSpontaneous\u003c/em\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 104px;\"\u003e\n \u003cp\u003e50.8% (0.500 - 0.515)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 94px;\"\u003e\n \u003cp\u003e70.8% (0.689 - 0.726)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 104px;\"\u003e\n \u003cp\u003e39.2% (0.383 - 0.400)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 94px;\"\u003e\n \u003cp\u003e56.5% (0.515 - 0.614)\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 161px;\"\u003e\n \u003cp\u003eFetal presentation\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 104px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 94px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 104px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 94px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 104px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 161px;\"\u003e\n \u003cp\u003e\u003cem\u003eVertex\u003c/em\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 104px;\"\u003e\n \u003cp\u003e97.5% (0.972 - 0.977)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 94px;\"\u003e\n \u003cp\u003e97.4% (0.967 - 0.980)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 104px;\"\u003e\n \u003cp\u003e98.4% (0.982 - 0.986)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 94px;\"\u003e\n \u003cp\u003e97.9% (0.959 - 0.989)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd rowspan=\"2\" valign=\"top\" style=\"width: 104px;\"\u003e\n \u003cp\u003e\u0026lt;0.0001\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 161px;\"\u003e\n \u003cp\u003e\u003cem\u003eOthers\u003c/em\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 104px;\"\u003e\n \u003cp\u003e2.5% (0.023 - 0.028)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 94px;\"\u003e\n \u003cp\u003e2.6% (0.020 - 0.033)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 104px;\"\u003e\n \u003cp\u003e1.6% (0.014 - 0.018)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 94px;\"\u003e\n \u003cp\u003e2.1% (0.011 - 0.041)\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 161px;\"\u003e\n \u003cp\u003eFetal Position\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 104px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 94px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 104px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 94px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 104px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 161px;\"\u003e\n \u003cp\u003e\u003cem\u003eNon-Vertex\u003c/em\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 104px;\"\u003e\n \u003cp\u003e2.5% (0.023 - 0.028)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 94px;\"\u003e\n \u003cp\u003e2.6% (0.020 - 0.033)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 104px;\"\u003e\n \u003cp\u003e1.6% (0.014 - 0.018)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 94px;\"\u003e\n \u003cp\u003e2.1% (0.011 - 0.041)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd rowspan=\"3\" valign=\"top\" style=\"width: 104px;\"\u003e\n \u003cp\u003e\u0026lt;0.0001\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 161px;\"\u003e\n \u003cp\u003e\u003cem\u003eVertex, occiput anterior\u003c/em\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 104px;\"\u003e\n \u003cp\u003e53.5% (0.528 - 0.543)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 94px;\"\u003e\n \u003cp\u003e46.4% (0.444 - 0.484)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 104px;\"\u003e\n \u003cp\u003e60.6% (0.597 - 0.614)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 94px;\"\u003e\n \u003cp\u003e52.4% (0.473 - 0.573)\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 161px;\"\u003e\n \u003cp\u003e\u003cem\u003eVertex, not occiput anterior\u003c/em\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 104px;\"\u003e\n \u003cp\u003e43.9% (0.432 - 0.447)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 94px;\"\u003e\n \u003cp\u003e51.0% (0.490 - 0.530)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 104px;\"\u003e\n \u003cp\u003e37.8% (0.370 - 0.387)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 94px;\"\u003e\n \u003cp\u003e45.5% (0.406 - 0.506)\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 161px;\"\u003e\n \u003cp\u003eMaternal position\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 104px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 94px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 104px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 94px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 104px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 161px;\"\u003e\n \u003cp\u003e\u003cem\u003eDorsal/lithotomy/semi recumbent\u003c/em\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 104px;\"\u003e\n \u003cp\u003e90.1% (0.896 - 0.905)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 94px;\"\u003e\n \u003cp\u003e79.6% (0.779 - 0.811)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 104px;\"\u003e\n \u003cp\u003e92.7% (0.922 - 0.931)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 94px;\"\u003e\n \u003cp\u003e88.7% (0.852 - 0.915)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd rowspan=\"4\" valign=\"top\" style=\"width: 104px;\"\u003e\n \u003cp\u003e\u0026lt;0.0001\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 161px;\"\u003e\n \u003cp\u003e\u003cem\u003eBirth stool/squatting/Standing/Water\u003c/em\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 104px;\"\u003e\n \u003cp\u003e1.3% (0.012 - 0.015)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 94px;\"\u003e\n \u003cp\u003e2.5% (0.019 - 0.032)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 104px;\"\u003e\n \u003cp\u003e0.7% (0.006 - 0.009)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 94px;\"\u003e\n \u003cp\u003e0.8% (0.003 - 0.023)\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 161px;\"\u003e\n \u003cp\u003e\u003cem\u003e\u0026nbsp;All fours/kneeling\u003c/em\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 104px;\"\u003e\n \u003cp\u003e5.2% (0.049 - 0.056)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 94px;\"\u003e\n \u003cp\u003e12.1% (0.108 - 0.135)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 104px;\"\u003e\n \u003cp\u003e4.1% (0.038 - 0.045)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 94px;\"\u003e\n \u003cp\u003e6.5% (0.045 - 0.095)\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 161px;\"\u003e\n \u003cp\u003e\u003cem\u003eLateral\u003c/em\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 104px;\"\u003e\n \u003cp\u003e3.4% (0.031 - 0.036)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 94px;\"\u003e\n \u003cp\u003e5.9% (0.050 - 0.069)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 104px;\"\u003e\n \u003cp\u003e2.5% (0.023 - 0.028)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 94px;\"\u003e\n \u003cp\u003e3.9% (0.024 - 0.064)\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 161px;\"\u003e\n \u003cp\u003eRegional Anesthesia use in labor\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 104px;\"\u003e\n \u003cp\u003e42.9% (0.422 - 0.437)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 94px;\"\u003e\n \u003cp\u003e13.2% (0.119 - 0.146)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 104px;\"\u003e\n \u003cp\u003e51.9% (0.510 - 0.528)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 94px;\"\u003e\n \u003cp\u003e24.6% (0.206 - 0.292)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 104px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 161px;\"\u003e\n \u003cp\u003eDuration of labor\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 104px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 94px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 104px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 94px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 104px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 161px;\"\u003e\n \u003cp\u003e\u003cem\u003eTotal\u0026nbsp;\u003c/em\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 104px;\"\u003e\n \u003cp\u003e4.97 (2.70-8.35)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 94px;\"\u003e\n \u003cp\u003e3.80 (2.22-6.23)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 104px;\"\u003e\n \u003cp\u003e5.47 (3.20-8.53)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 94px;\"\u003e\n \u003cp\u003e3.62 (2.02-6.40)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd rowspan=\"3\" valign=\"top\" style=\"width: 104px;\"\u003e\n \u003cp\u003e\u0026lt;0.0001\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 161px;\"\u003e\n \u003cp\u003e\u003cem\u003eFirst stage\u003c/em\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 104px;\"\u003e\n \u003cp\u003e4.17 (2.25-7.25)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 94px;\"\u003e\n \u003cp\u003e3.33 (1.88-5.50)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 104px;\"\u003e\n \u003cp\u003e4.58 (2.58-7.50)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 94px;\"\u003e\n \u003cp\u003e3.12 (1.67-5.66)\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 161px;\"\u003e\n \u003cp\u003e\u003cem\u003eSecond stage\u003c/em\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 104px;\"\u003e\n \u003cp\u003e0.47 (0.18-1.20)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 94px;\"\u003e\n \u003cp\u003e0.35 (0.18-0.68)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 104px;\"\u003e\n \u003cp\u003e0.50 (0.17-1.22)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 94px;\"\u003e\n \u003cp\u003e0.33 (0.17-0.58)\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 161px;\"\u003e\n \u003cp\u003eIntrapartum passage of meconium\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 104px;\"\u003e\n \u003cp\u003e12.9% (0.125 - 0.135)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 94px;\"\u003e\n \u003cp\u003e11.2% (0.100 - 0.125)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 104px;\"\u003e\n \u003cp\u003e14.5% (0.139 - 0.152)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 94px;\"\u003e\n \u003cp\u003e17.3% (0.138 - 0.214)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 104px;\"\u003e\n \u003cp\u003e\u0026lt;0.0001\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 161px;\"\u003e\n \u003cp\u003eSentinel events\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 104px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 94px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 104px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 94px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 104px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 161px;\"\u003e\n \u003cp\u003e\u003cem\u003eCord accident\u003c/em\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 104px;\"\u003e\n \u003cp\u003e16.5% (0.160 - 0.171)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 94px;\"\u003e\n \u003cp\u003e17.5% (0.160 - 0.191)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 104px;\"\u003e\n \u003cp\u003e15.4% (0.148 - 0.160)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 94px;\"\u003e\n \u003cp\u003e16.5% (0.131 - 0.205)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd rowspan=\"2\" valign=\"top\" style=\"width: 104px;\"\u003e\n \u003cp\u003e0.07\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 161px;\"\u003e\n \u003cp\u003e\u003cem\u003eShoulder dystocia/abruption/failed instrumental\u003c/em\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 104px;\"\u003e\n \u003cp\u003e1.2% (0.011 - 0.014)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 94px;\"\u003e\n \u003cp\u003e1.4% (0.010 - 0.019)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 104px;\"\u003e\n \u003cp\u003e1.2% (0.010 - 0.014)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 94px;\"\u003e\n \u003cp\u003e1.8% (0.009 - 0.037)\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 161px;\"\u003e\n \u003cp\u003eMode of birth\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 104px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 94px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 104px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 94px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 104px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 161px;\"\u003e\n \u003cp\u003e\u003cem\u003eVaginal\u003c/em\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 104px;\"\u003e\n \u003cp\u003e70.4% (0.697 - 0.711)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 94px;\"\u003e\n \u003cp\u003e90.1% (0.888 - 0.912)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 104px;\"\u003e\n \u003cp\u003e52.1% (0.512 - 0.530)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 94px;\"\u003e\n \u003cp\u003e67.3% (0.624 - 0.718)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd rowspan=\"4\" valign=\"top\" style=\"width: 104px;\"\u003e\n \u003cp\u003e0.00\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 161px;\"\u003e\n \u003cp\u003e\u003cem\u003eForceps assisted\u003c/em\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 104px;\"\u003e\n \u003cp\u003e10.9% (0.105 - 0.114)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 94px;\"\u003e\n \u003cp\u003e2.5% (0.019 - 0.032)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 104px;\"\u003e\n \u003cp\u003e16.5% (0.159 - 0.172)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 94px;\"\u003e\n \u003cp\u003e8.9% (0.064 - 0.122)\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 161px;\"\u003e\n \u003cp\u003e\u003cem\u003eVacuum assisted\u003c/em\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 104px;\"\u003e\n \u003cp\u003e8.4% (0.080 - 0.088)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 94px;\"\u003e\n \u003cp\u003e4.4% (0.036 - 0.053)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 104px;\"\u003e\n \u003cp\u003e13.8% (0.132 - 0.144)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 94px;\"\u003e\n \u003cp\u003e17.5% (0.141 - 0.217)\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 161px;\"\u003e\n \u003cp\u003eCS\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 104px;\"\u003e\n \u003cp\u003e10.3% (0.099 - 0.108)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 94px;\"\u003e\n \u003cp\u003e3.0% (0.024 - 0.038)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 104px;\"\u003e\n \u003cp\u003e17.6% (0.169 - 0.182)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 94px;\"\u003e\n \u003cp\u003e6.3% (0.043 - 0.092)\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 161px;\"\u003e\n \u003cp\u003eIncidence of asphyxia*\u0026nbsp;\u0026nbsp;(number of records, % proportion,95%CI)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 104px;\"\u003e\n \u003cp\u003e383,\u003c/p\u003e\n \u003cp\u003e2.2 (1.98 - 2.41)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 94px;\"\u003e\n \u003cp\u003e57,\u003c/p\u003e\n \u003cp\u003e2.4% (1.87 - 3.11)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 104px;\"\u003e\n \u003cp\u003e400,\u003c/p\u003e\n \u003cp\u003e3.3% (3.03 - 3.68)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 94px;\"\u003e\n \u003cp\u003e20,\u003c/p\u003e\n \u003cp\u003e5.2 (3.41 - 7.94)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 104px;\"\u003e\n \u003cp\u003e\u0026lt;0.0001\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003c/tbody\u003e\n\u003c/table\u003e\n\u003cp\u003e\u003cem\u003eData are median (IQR)(as not normally distributed), n (%,95% Confidence Interval).\u003c/em\u003e\u003c/p\u003e\n\u003cp\u003e\u003cem\u003eOnly BMI had missing data as described\u0026nbsp;\u003c/em\u003e\u003c/p\u003e\n\u003cp\u003e* \u0026nbsp;By composite definition; \u0026dagger; e.g. Abruption, post-term; \u0026sect; Including hypertensive disease in pregnancy; \u0026dagger;\u0026dagger; e.g. macrosomia, polyhydramnios, unstable lie\u003cstrong\u003e\u003c/strong\u003e\u003cstrong\u003e\u0026nbsp;\u003c/strong\u003e\u003c/p\u003e\u003cp\u003eThis table summarizes the clinical, obstetric, and neonatal characteristics of training cohort women stratified by fetal heart rate monitoring modality (ultrasound or ECG) and signal dropout level (\u0026gt;30% or \u0026le;30%). Data are presented as medians (IQR) or proportions with 95% confidence intervals. Differences across groups were assessed using appropriate statistical tests.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eTable 2: Causal Effects of Mode-Specific Signal Dropout on Perinatal Asphyxia in Training and Validation Cohorts\u003c/strong\u003e\u003c/p\u003e\n\u003ctable border=\"0\" cellspacing=\"0\" cellpadding=\"0\" width=\"616\"\u003e\n \u003ctbody\u003e\n \u003ctr\u003e\n \u003ctd valign=\"bottom\" style=\"width: 217px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eDropout Group\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 51px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eModel\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 103px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eAdjusted Odds Ratio (aOR)\u003csup\u003ea\u003c/sup\u003e\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 94px;\"\u003e\n \u003cp\u003e\u003cstrong\u003e95% CI\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 151px;\"\u003e\n \u003cp\u003e\u003cstrong\u003ep-value\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd colspan=\"5\" valign=\"bottom\" style=\"width: 616px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eTraining dataset\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"bottom\" style=\"width: 217px;\"\u003e\n \u003cp\u003eECG Low Dropout\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 51px;\"\u003e\n \u003cp\u003eIPW\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 103px;\"\u003e\n \u003cp\u003e1.51\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 94px;\"\u003e\n \u003cp\u003e1.32-1.71\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 151px;\"\u003e\n \u003cp\u003e\u0026lt;0.001\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"bottom\" style=\"width: 217px;\"\u003e\n \u003cp\u003eUS High Dropout\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 51px;\"\u003e\n \u003cp\u003e\u003cstrong\u003e\u0026nbsp;\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 103px;\"\u003e\n \u003cp\u003e1.25\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 94px;\"\u003e\n \u003cp\u003e0.99-1.58\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 151px;\"\u003e\n \u003cp\u003e0.06\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"bottom\" style=\"width: 217px;\"\u003e\n \u003cp\u003eECG High Dropout\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 51px;\"\u003e\n \u003cp\u003e\u003cstrong\u003e\u0026nbsp;\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 103px;\"\u003e\n \u003cp\u003e3.14\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 94px;\"\u003e\n \u003cp\u003e2.20-4.48\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 151px;\"\u003e\n \u003cp\u003e\u0026lt;0.001\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"bottom\" style=\"width: 217px;\"\u003e\n \u003cp\u003eECG High Dropout at threshold 5%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 51px;\"\u003e\n \u003cp\u003eTMLE\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 103px;\"\u003e\n \u003cp\u003e1.22\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 94px;\"\u003e\n \u003cp\u003e0.84-1.78\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 151px;\"\u003e\n \u003cp\u003e0.29\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 217px;\"\u003e\n \u003cp\u003eECG High Dropout at threshold 10%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd rowspan=\"5\" valign=\"bottom\" style=\"width: 51px;\"\u003e\n \u003cp\u003e\u003cstrong\u003e\u0026nbsp;\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 103px;\"\u003e\n \u003cp\u003e1.19\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 94px;\"\u003e\n \u003cp\u003e0.94-1.51\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 151px;\"\u003e\n \u003cp\u003e0.15\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 217px;\"\u003e\n \u003cp\u003eECG High Dropout at threshold 15%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 103px;\"\u003e\n \u003cp\u003e1.21\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 94px;\"\u003e\n \u003cp\u003e0.94-1.55\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 151px;\"\u003e\n \u003cp\u003e0.14\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 217px;\"\u003e\n \u003cp\u003eECG High Dropout at threshold 20%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 103px;\"\u003e\n \u003cp\u003e1.56\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 94px;\"\u003e\n \u003cp\u003e1.16-2.09\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 151px;\"\u003e\n \u003cp\u003e0.00\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 217px;\"\u003e\n \u003cp\u003eECG High Dropout at threshold 25%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 103px;\"\u003e\n \u003cp\u003e1.74\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 94px;\"\u003e\n \u003cp\u003e1.18-2.56\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 151px;\"\u003e\n \u003cp\u003e0.00\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 217px;\"\u003e\n \u003cp\u003eECG High Dropout at threshold 30%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 103px;\"\u003e\n \u003cp\u003e2.08\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 94px;\"\u003e\n \u003cp\u003e1.21-3.58\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 151px;\"\u003e\n \u003cp\u003e0.01\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd colspan=\"5\" valign=\"bottom\" style=\"width: 616px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eExternal Validation Dataset\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd colspan=\"5\" valign=\"bottom\" style=\"width: 616px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eCzech dataset\u0026nbsp;\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"bottom\" style=\"width: 217px;\"\u003e\n \u003cp\u003eECG High Dropout\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 51px;\"\u003e\n \u003cp\u003eIPW\u003csup\u003eb\u003c/sup\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 103px;\"\u003e\n \u003cp\u003e4.02\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 94px;\"\u003e\n \u003cp\u003e1.41-11.44\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 151px;\"\u003e\n \u003cp\u003e0.01\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"bottom\" style=\"width: 217px;\"\u003e\n \u003cp\u003eECG Low Dropout\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd rowspan=\"2\" valign=\"bottom\" style=\"width: 51px;\"\u003e\u003cbr\u003e\u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 103px;\"\u003e\n \u003cp\u003e0.24\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 94px;\"\u003e\n \u003cp\u003e0.06-0.96\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 151px;\"\u003e\n \u003cp\u003e0.04\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"bottom\" style=\"width: 217px;\"\u003e\n \u003cp\u003eUS High Dropout\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 103px;\"\u003e\n \u003cp\u003e1.90\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 94px;\"\u003e\n \u003cp\u003e0.94-3.82\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 151px;\"\u003e\n \u003cp\u003e0.07\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"bottom\" style=\"width: 217px;\"\u003e\n \u003cp\u003eECG High Dropout at threshold 30%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 51px;\"\u003e\n \u003cp\u003eTMLE\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 103px;\"\u003e\n \u003cp\u003e2.83\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 94px;\"\u003e\n \u003cp\u003e2.13-16.36\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 151px;\"\u003e\n \u003cp\u003e0.01\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd colspan=\"5\" valign=\"bottom\" style=\"width: 616px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eAustralian dataset 2 (Werribee)\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"bottom\" style=\"width: 217px;\"\u003e\n \u003cp\u003eECG High Dropout\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 51px;\"\u003e\n \u003cp\u003eIPW\u003csup\u003eb\u003c/sup\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 103px;\"\u003e\n \u003cp\u003e2.22\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 94px;\"\u003e\n \u003cp\u003e1.61-3.05\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 151px;\"\u003e\n \u003cp\u003e\u0026lt;0.001\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"bottom\" style=\"width: 217px;\"\u003e\n \u003cp\u003eECG Low Dropout\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd rowspan=\"2\" valign=\"bottom\" style=\"width: 51px;\"\u003e\u003cbr\u003e\u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 103px;\"\u003e\n \u003cp\u003e1.57\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 94px;\"\u003e\n \u003cp\u003e1.31-1.89\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 151px;\"\u003e\n \u003cp\u003e\u0026lt;0.001\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"bottom\" style=\"width: 217px;\"\u003e\n \u003cp\u003eUS High Dropout\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 103px;\"\u003e\n \u003cp\u003e0.69\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 94px;\"\u003e\n \u003cp\u003e0.33-1.46\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 151px;\"\u003e\n \u003cp\u003e0.33\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"bottom\" style=\"width: 217px;\"\u003e\n \u003cp\u003eECG High Dropout at threshold 30%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 51px;\"\u003e\n \u003cp\u003eTMLE\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 103px;\"\u003e\n \u003cp\u003e1.91\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 94px;\"\u003e\n \u003cp\u003e1.11-3.28\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 151px;\"\u003e\n \u003cp\u003e0.02\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003c/tbody\u003e\n\u003c/table\u003e\n\u003cp\u003e\u003csup\u003ea\u0026nbsp;\u003c/sup\u003e\u003cem\u003eAll comparisons are made against the Ultrasound Low Dropout group, which serves as the reference category in both IPW and TMLE models.\u003c/em\u003e\u003c/p\u003e\n\u003cp\u003e\u003csup\u003eb\u0026nbsp;\u003c/sup\u003e\u003cem\u003eAll IPW modelling\u0026nbsp;\u003c/em\u003e\u003cem\u003ewas on a 30% dropout cutoff\u003c/em\u003e\u003c/p\u003e\u003cp\u003eAdjusted odds ratios (aORs) from inverse probability\u0026ndash;weighted logistic regression and dose-dependent targeted maximum likelihood estimation (TMLE) analyses show the association between mode-specific signal dropout and perinatal asphyxia in training and validation cohorts.\u0026nbsp;\u003c/p\u003e"}],"fulltextSource":"","fullText":"","funders":[],"hasAdminPriorityOnWorkflow":false,"hasManuscriptDocX":true,"hasOptedInToPreprint":true,"hasPassedJournalQc":"","hasAnyPriority":false,"hideJournal":true,"highlight":"","institution":"","isAcceptedByJournal":false,"isAuthorSuppliedPdf":false,"isDeskRejected":"","isHiddenFromSearch":false,"isInQc":false,"isInWorkflow":false,"isPdf":false,"isPdfUpToDate":true,"isWithdrawnOrRetracted":false,"journal":{"display":true,"email":"[email protected]","identity":"researchsquare","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":true,"externalIdentity":"","sideBox":"","snPcode":"","submissionUrl":"/submission","title":"Research Square","twitterHandle":"researchsquare","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"","reportingPortfolio":"","inReviewEnabled":false,"inReviewRevisionsEnabled":true},"keywords":"","lastPublishedDoi":"10.21203/rs.3.rs-7122719/v1","lastPublishedDoiUrl":"https://doi.org/10.21203/rs.3.rs-7122719/v1","license":{"name":"CC BY 4.0","url":"https://creativecommons.org/licenses/by/4.0/"},"manuscriptAbstract":"Cardiotocography (CTG) is the cornerstone of intrapartum surveillance, yet its predictive value for perinatal asphyxia remains limited. Modern monitors record fetal heart rate (FHR) via Doppler ultrasound or intrapartum fetal electrocardiography (ECG), the latter usable only after cervical dilatation. We analysed 52,757 labours ≥36 weeks’ gestation across three hospitals—two in Australia and one in Europe—treating mode-specific FHR dropout (zero-value samples signifying missing valid biosignal) as an informative exposure rather than benign artifact. High dropout (\u003e30% of the CTG recording available for analysis) was evaluated using doubly robust causal machine learning and survival analysis, incorporating inverse-probability weighting, targeted maximum-likelihood estimation (TMLE), and weighted Cox models. In the training cohort (n = 36,792), high ECG-mode dropout tripled the odds of perinatal asphyxia (OR 3.14, 95% CI: 2.20–4.48) and quadrupled the hazard (HR 4.76, 95% CI: 2.98–7.62) compared to ultrasound-mode low dropout. In crude analyses, this corresponded to an absolute risk increase of 5.0% (1 in 20 births). TMLE-adjusted models confirmed a 2.27% absolute risk increase—equivalent to one additional case per 44 labor episodes (95% CI: 28–83). External validation confirmed consistent associations in European (OR 4.02) and Australian (OR 2.22) cohorts. “Imperfect” ECG signals thus offer biologically plausible and trustworthy inputs for standalone or ensemble artificial intelligence models, enabling transparent, mode-aware decision support in real-time fetal monitoring.","manuscriptTitle":"Multimodal Causal Machine Learning for Fetal Asphyxia Risk Prediction from Imperfect Monitoring Signals","msid":"","msnumber":"","nonDraftVersions":[{"code":1,"date":"2025-08-13 05:53:29","doi":"10.21203/rs.3.rs-7122719/v1","editorialEvents":[{"type":"communityComments","content":0}],"status":"published","journal":{"display":true,"email":"[email protected]","identity":"researchsquare","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":true,"externalIdentity":"","sideBox":"","snPcode":"","submissionUrl":"/submission","title":"Research Square","twitterHandle":"researchsquare","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"","reportingPortfolio":"","inReviewEnabled":false,"inReviewRevisionsEnabled":true}}],"origin":"","ownerIdentity":"707018e0-6048-4cec-bc62-dcbce5c41ac4","owner":[],"postedDate":"August 13th, 2025","published":true,"recentEditorialEvents":[],"rejectedJournal":[],"revision":"","amendment":"","status":"posted","subjectAreas":[{"id":52781375,"name":"Health sciences/Cardiology"},{"id":52781376,"name":"Health sciences/Health care"},{"id":52781377,"name":"Health sciences/Medical research"}],"tags":[],"updatedAt":"2025-11-16T11:38:18+00:00","versionOfRecord":[],"versionCreatedAt":"2025-08-13 05:53:29","video":"","vorDoi":"","vorDoiUrl":"","workflowStages":[]},"version":"v1","identity":"rs-7122719","journalConfig":"researchsquare"},"__N_SSP":true},"page":"/article/[identity]/[[...version]]","query":{"redirect":"/article/rs-7122719","identity":"rs-7122719","version":["v1"]},"buildId":"8U1c8b4HqxoKbykW_rLl7","isFallback":false,"isExperimentalCompile":false,"dynamicIds":[84888],"gssp":true,"scriptLoader":[]}