Dual-output, web-based risk assessment system for cesarean section due to dystocia: integration of logistic regression and risk scoring models
Article information
Abstract
Objective
To develop a web-based risk assessment system to predict cesarean section (CS) due to dystocia at admission in nulliparous term singleton vertex pregnancies, tailored for Korean women.
Methods
This case-control study analyzed the data of 126 women with CS due to dystocia and 490 women who had vaginal deliveries. Eight predictors-gestational age, maternal age, maternal height, pre-gestational body mass index, birth weight, fetal sex, cervical dilatation at admission, and maternal-fetal ratio-were identified using multivariate logistic regression. The system integrated both logistic regression and risk-scoring models simultaneously to provide individualized risk probabilities and categorical risk levels.
Results
The model demonstrated strong predictive accuracy, with an area under the receiver operating characteristic curve of 0.86. Risk stratification classified the patients into low-, intermediate-, and high-risk groups, corresponding to CS rates of 1.6, 47.6, and 50.8%, respectively (P<0.001).
Conclusion
This dual-output, user-friendly, and admission-based web system enhances interpretability and supports personalized counseling and evidence-based decision-making. Specifically designed for Korean women, it enables the early identification of high-risk cases and may help reduce unnecessary operative interventions. Therefore, further multicenter studies are warranted.
Introduction
Cesarean section (CS) is a vital obstetric intervention, but is often overused, particularly in low-risk pregnancies. In Korea, the CS rate has steadily increased from 36.7% in 2012 to 58.7% in 2023 [1,2], far exceeding the World Health Organization’s recommended range of 10–15% [3]. Maternal requests without medical indications account for 6–7% of CS cases in Korea [4,5].
Recent nationwide analyses have demonstrated a continuous increase in CS rates and evolving maternal risk profiles in South Korea, emphasizing the need for updated and clinically applicable risk assessment tools tailored to Korean women [6].
Although national perinatal mortality has gradually declined, high-risk births continue to impose a substantial clinical burden, reinforcing the importance of optimizing intrapartum risk stratification [7].
Accurate tools to estimate CS risk are indispensable for guiding clinical decision-making, enhancing patient counseling, and minimizing unnecessary interventions. Recently, programmable tools, such as calculators, have emerged, offering care providers a straightforward and accessible method to convey individualized risk assessments. Since dystocia remains the primary indication for CS, there has been significant interest in developing reliable, programmable tools specifically designed to estimate CS risk in such cases. Contemporary tools, ranging from traditional logistic regression models [8–10] to advanced machine learning (ML)-based models [11–13], have been incorporated into obstetric practice to address these critical needs, enhance predictive accuracy, and support informed clinical decision-making. However, these tools often lack specific focus on dystocia [8,9], rely on variables unavailable at the time of admission [10], and require substantial computational resources [11–13]. Moreover, they have primarily been developed and validated in Western populations, limiting their applicability to Korean women [8–12].
Maternal anthropometric characteristics, particularly among Asian populations, have been shown to significantly influence fetal growth and perinatal outcomes, supporting the relevance of body size parameters such as body mass index (BMI), maternal height, and proportional indices in risk prediction [14].
This study introduces a web-based system tailored to predict dystocia-related CS in Korean women upon admission. By integrating logistic regression and risk-scoring models, the system delivers a predictive accuracy comparable to that of ML models while reducing computational complexity.
Materials and methods
1. Study design and data collection
Clinical data were retrospectively collected from the delivery registry of the Eulji University Hospital, covering the period from January 2014 to December 2017. The initial dataset included 2,079 women who delivered at 37 or more weeks of gestation with a cephalic presentation during nulliparous term singleton vertex (NTSV) pregnancies, extracted from the delivery database at the Eulji University Hospital, forming the dataset for this analysis. The exclusion criteria encompassed multiparous women, multiple gestations, non-vertex presentations, preterm labor, high-risk pregnancies, and the presence of one or more maternal or fetal conditions, including abnormal fetal heart rate tracing, vaginal delivery within 2 hours of admission, pre-labor CS, diabetes, gestational diabetes, pregnancy-induced hypertension, intrauterine growth restriction, placenta previa, placental abruption, active genital herpes, or other significant medical complications. Women with incomplete medical records or missing data on key variables were also excluded. After applying these criteria, 616 women were included in the final analysis, comprising 126 cases of CS due to dystocia and 490 controls who delivered vaginally.
The primary outcome was CS due to dystocia, defined as an average cervical dilatation rate of less than 0.5 cm per hour over a 2–4-hours period during the active phase of labor. If both dystocia and fetal distress were present, the patient was categorized as having CS due to dystocia. Clinical data included maternal age, height, prepregnancy BMI, and gestational age. Cervical dilatation at admission, fetal sex, estimated fetal weight (EFW) (assessed via ultrasonography within 24 hours of hospital admission), and symphysis-fundal height were also included. The maternal-fetal ratio (MFR) was calculated by dividing maternal height by symphysis-fundal height [15–17]. The active phase of labor was retrospectively defined according to guidelines established by the National Institute for Health and Care Excellence [18] and the American College of Obstetricians and Gynecologists (ACOG)/Society for Materna-Fetal Medicine (SMFM) [19]. This phase was defined as cervical dilatation of at least 4 cm with progressive changes, or 6 cm regardless of progression. Dystocia was defined according to the 2014 ACOG and SMFM criteria. Active-phase arrest was defined as no cervical change for ≥4 hours despite adequate uterine contractions or ≥6 hours with oxytocin augmentation after the onset of the active phase. Second-stage arrest was defined as no fetal descent after ≥3 hours of pushing in nulliparous women despite adequate uterine activity [19].
2. Statistical analysis
The first step in our analytical approach was to identify candidate variables for CS due to dystocia using univariate and multivariate logistic regression analyses. Potential variables were selected and categorized into three levels, excluding fetal sex, based on a literature review and the study team’s clinical judgment. The second step involved the construction of logistic regression and risk-scoring models. Potential predictors were selected via univariate analysis (P<0.10) and the final model was built using backward variable selection based on logistic regression analysis, requiring a P-value <0.05 for final inclusion. The predictive variables identified in the multivariate logistic regression were transformed into a risk score with weighted coefficients, converted into item scores, and summed to obtain a total score. The third step measured the area under the ROC curve (AUC) and calculated sensitivity, specificity, positive predictive value (PPV), negative predictive value (NPV), and likelihood ratios ([LR]+ and LR−). The fourth step evaluated model accuracy using a calibration curve and the Hosmer-Lemeshow goodness-of-fit test, supplemented by Akaike information criterion (AIC) and Bayesian information criterion (BIC) metrics, to assess model fit, where lower values indicated a better fit. Internal validation was performed by bootstrapping (200 replicates).
All statistical analyses were conducted using STATA version 13.1 (StataCorp., College Station, TX, USA) and calibration plots were generated using R version 3.2.3 (R Foundation for Statistical Computing, Vienna, Austria). G*Power version 3.1 (Heinrich Heine University Düsseldorf, Düsseldorf, Germany) was used to calculate statistical power, confirming that the sample size provided a post-hoc power of 88.4% with a two-sided α of 0.05.
Results
1. Patient characteristics and model development
The medical records of 616 patients were reviewed, of whom 20.5% (126) had undergone CS for dystocia during NTSV pregnancy. The descriptive characteristics of the patients are summarized in Table 1.
Obstetric characteristics of women with cesarean section due to dystocia (cases: n=126) and by vaginal delivery (controls: n=490)
The logistic regression model identified eight significant predictors of CS due to dystocia: gestational age, maternal age, maternal height, pre-gestational BMI, EFW (assessed via ultrasonography within 24 hours of hospital admission), fetal sex, cervical dilatation at admission, and MFR. The predictors are listed in Table 2. The final logistic regression equation for calculating the probability of CS is presented in Supplementary Material 1.
Logistic regression analysis of risk indicators for cesarean section due to dystocia
To develop a risk-scoring model, each predictor in the logistic regression model was assigned a specific score derived from the regression coefficient and then transformed to 1 by dividing the smallest coefficient in the pre-pregnancy BMI (0.32). These scores were converted to final scores by rounding up (see Supplementary Table 1). The risk-scoring model had a total score ranging from 0 to 8. The total risk score was then categorized into low-(<12), intermediate-(12–21.9), and high-risk (≥22) groups. Thresholds were empirically derived through stratified cross-tabulation analyses of the total scores and delivery outcomes. As shown in Supplementary Table 2, the cutoff range of 12–21.9 provided the highest discrimination (χ2=180.9; ρ=0.475), demonstrating optimal separation among risk groups and the most distinct LR for cesarean delivery due to dystocia. Risk stratification categorized patients into low-, intermediate-, and high-risk groups, with CS rates of 1.6, 47.6, and 50.8%, respectively. In comparison, among the controls, the distributions were 31.0, 63.7, and 5.3% in the low-, intermediate-, and high-risk groups, respectively. These percentages highlight significant risk stratification, with a higher concentration of high-risk patients in the CS group, underscoring the model’s ability to effectively distinguish risk categories. The LR for the CS with low-, intermediate-, and high-risk groups was 0.05 (95% confidence interval [CI], 0.01–0.20), 0.75 (95% CI, 0.62–0.91), and 9.57 (95% CI, 6.34–14.45), respectively. Women in the low-risk group were only 0.05 times as likely to undergo CS, whereas those in the high-risk group were 9.57 times more likely, demonstrating a significant increase in risk for the high-risk category (Table 3).
Distribution of risk levels among cases and controls
2. Model performance
The logistic regression model demonstrated strong predictive performance, with an AUC of 0.8583 (95% CI, 0.8217–0.8949) derived from receiver operating characteristic curve analysis. Similarly, the risk-scoring model achieved an AUC of 0.8580 (95% CI, 0.8214–0.8945), indicating comparable discriminatory ability (Fig. 1). Calibration plots for the logistic regression model showed good alignment between predicted and observed probabilities, with a Hosmer-Lemeshow test P-value of 0.20 (Fig. 2). Additional metrics, including AIC and BIC values (logistic regression: AIC=457.8; BIC=528.6; risk-scoring model: AIC=444.0; BIC=483.9), confirmed robust calibration. The comparison of AUCs revealed no significant differences between the models (P=0.72), suggesting a comparable performance for predicting CS risk due to dystocia.
Receiver operating characteristic curve comparison between the logistic regression model and risk scoring model in the development sets. AUC, area under the receiver operating characteristic curve.
Performance of the logistic regression model, the calibration plots comparing the actual probabilities (y) and predicted probabilities (x) of the use of the logistic regression model to predict CS due to dystocia. CS, cesarean section.
3. Cut-off points and diagnostic accuracy
At the logistic regression model cutoff point of 0.5, the sensitivity and specificity were 95.1% and 47.6%, respectively, with a PPV of 87.6% and a NPV of 71.4%. For the high-risk group (score ≥22) in the risk scoring model, the PPV and NPV were 71.1% and 88.2%, respectively, with an LR+ of 9.57 and an LR− of 0.52 (Tables 4, 5).
Diagnostic performance of logistic regression model by cutoff probability
Predictive performance by total risk-score levels
4. Internal validation
Internal validation using a bootstrapping (200 replicates) revealed minimal optimism with adjusted AUCs of 0.846 and 0.851 for the logistic regression and risk-scoring models, respectively. This confirmed robust internal validity, as all retained parameters demonstrated statistical significance (P<0.05) in 90% of the bootstrap samples.
5. Dual-output, web-based risk assessment system
An integrated web interface was developed in which the logistic and risk-scoring models operate simultaneously and complementarily. The logistic model provides a quantitative risk probability, whereas the risk-scoring model categorizes patients as low, intermediate, or high risk. The results are displayed together to aid clinical decision-making and patient counseling, thereby enhancing interpretability and practical usability. This dual-output web-based risk assessment tool is freely available online at https://rpzp8w.csb.app (Fig. 3).
Dual-output, web-based assessment system for predicting CS due to dystocia at admission (cesarean section prediction for dystocia). BMI, body mass index; NTSV, nulliparous term singleton vertex; CS, cesarean section.
Discussion
This study developed and validated a dual-output, web-based risk assessment system integrating logistic regression and risk-scoring models to predict CS due to dystocia in NTSV pregnancies. The system provides comprehensive risk probabilities and categorical risk levels and acts as a practical and user-friendly tool to support decision-making and patient counseling.
Methodological strategies to improve CS prediction have previously included the use of homogeneous study populations [20,21], adherence to standardized labor management protocols [22,23], and incorporation of clinically relevant admission-based variables to enhance early risk assessment and clinical applicability [24–26]. Building on this foundation, the present study focused on a homogeneous NTSV cohort that adhered to the ACOG/SMFM labor management guidelines and utilized admission-based variables routinely available in clinical practice. Previous admission-based models have identified maternal anthropometry, fetal size, and cervical findings as predictors of delivery outcome [27–29]. The strong associations observed among maternal height, MFR, and EFW underscore the central role of cephalopelvic proportionality in determining the likelihood of successful vaginal delivery. Shorter maternal stature and smaller MFR values indicate a relatively narrower pelvic capacity or larger fetal dimensions, which physiologically increases the risk of labor arrest, consistent with their elevated odds ratio (OR) (Tables 1, 2). Similarly, a larger EFW and male fetal sex, both of which are associated with increased soft tissue and bony resistance during descent, were significantly linked to dystocia-related cesarean delivery. Cervical dilatation of <3 cm at admission also demonstrated a strong association with cesarean delivery, reflecting a less favorable starting point and early inefficient labor progress. Together, these findings provide a clinical explanation for the OR patterns observed in Tables 1, 2, and illustrate why these predictors emerged as key determinants in the admission-based model. Building on this framework, the present study adopted a physiologically integrated approach linking maternal-fetal proportionality with cervical progress upon admission to enhance interpretability and clinical utility. Consequently, the developed model demonstrated robust discrimination and calibration (AUC, 0.86), supporting its clinical applicability as a practical admission-based prediction tool for dystocia-related cesarean deliveries. This performance compares favorably with previously reported admission-based models (AUC, 0.69–0.79) [10,29–31] and is comparable to more resource-intensive ML approaches (AUC, 0.82–0.84) [11,12]. By combining statistical rigor with bedside interpretability, our web-based implementation provides point-of-care usability without requiring a specialized infrastructure, supporting its feasibility in diverse obstetric environments. Within the web interface, the logistic regression and risk-scoring components operate simultaneously and complementarily, integrating quantitative precision and categorical interpretability. A total score ≥22 identified high-risk cases, while scores <12 indicated low risk, enabling rapid triage and individualized preparation for potential operative delivery. This dual-output configuration enhances clinician-patient communication and functions as a practical bedside counseling tool. Clinically, the low-, intermediate-, and high-risk groups showed a clear gradient in the observed probability of cesarean delivery due to dystocia; the low-risk group demonstrated the lowest cesarean proportion, the intermediate group showed a moderate risk, and the high-risk group demonstrated the highest probability. This stepwise pattern supports the practical use of this scoring system for simple risk stratification during admission to the hospital. Our web-based dual-output system is consistent with the emerging movement toward digital and artificial intelligence-assisted decision support tools in obstetrics while retaining the transparency and clinical interpretability of conventional regression-based models [32]. Although the specificity of the model was moderate (47.6%), the threshold was intentionally optimized for high sensitivity (95.1%) to prioritize patient safety and minimize missed cases of high-risk dystocia, which is clinically appropriate for admission-based risk assessment. This safety-oriented configuration ensures that high-risk patients are promptly identified at admission, enabling timely counseling and operative preparation.
While ultrasound-based EFW is commonly available in tertiary centers, resource-limited settings may rely on symphysis-fundal height or Leopold’s maneuvers [33–36]. Anthropometric ratios such as the MFR [15–17] and head circumference-to-maternal height ratio [37] complement EFW by representing cephalopelvic proportionality. Although MFR is not yet universally applied and may vary among operators, it provides a physiologically meaningful surrogate for cephalopelvic balance and is consistently measurable within standardized admission assessments. Future studies should evaluate simplified indices to facilitate their broader adoption. Conceptually, the MFR reflects the cephalopelvic proportionality that underlies labor dystocia, serving as a simplified anthropometric approximation of the maternal-fetal relationship. As maternal and fundal height distributions vary across Asian populations, incorporating MFR may provide population-specific predictive values beyond those of traditional fetal size parameters.
The present model was intentionally designed as a static admission-based tool for early triage and counseling. Dynamic intrapartum variables, such as uterine activity, cervical change rate, and fetal descent, were excluded because continuous monitoring was beyond the scope of this study. Nevertheless, recent research has demonstrated the potential for integrating dynamic labor phase parameters [38–40]. Future work will extend this framework to an adaptive, phase-specific predictive model that incorporates real-time labor metrics.
This study has some inherent limitations. The study was conducted at a single center using a retrospective design, which may have introduced selection and information biases. Additionally, because no external validation cohort was available, the generalizability of the model beyond the study population remains limited. Future prospective multicenter studies are needed to confirm reproducibility and strengthen external validation. Cervical dilatation at admission is also inherently susceptible to interobserver variation and timing-related differences. Although standardized examination protocols were used at our institution, such variability cannot be eliminated and may have influenced the observed association with dystocia-related cesarean delivery. Beyond the methodological limitations, practical implementation warrants further consideration. Although the web-based system was technically validated, formal usability testing involving end users (clinicians and patients) was not performed, and the data security or sustainability of the online platform was not assessed. Similarly, while the system was developed to improve workflow efficiency and accessibility, no formal cost-benefit or time-efficiency analysis was conducted. Future research should incorporate structured usability evaluation, compliance with healthcare data protection, and systematic cost-effectiveness assessments to ensure sustainable clinical integration.
In conclusion, this study developed an integrative, dual-output, web-based risk assessment system to predict cesarean delivery due to dystocia upon hospital admission. By combining maternal, fetal, and cervical parameters within a unified physiological model, the system achieved strong discrimination and calibration (AUC=0.86), demonstrating predictive performance within the upper range of previously published models, including recent ML-based approaches [9,10] while maintaining superior interpretability and accessibility. Dual logistic and scoring outputs enhance transparency and clinical usability, supporting shared decision-making in obstetric care. Future multicenter prospective studies incorporating dynamic labor parameters are warranted to refine, validate, and expand this framework for broader clinical applications.
Supplementary Information
Notes
Conflict of interest
The authors declare no conflicts of interest.
Ethical approval
The study protocol was reviewed and approved by the Institutional Review Board of Eulji University Hospital in 2017 (reference number: 2017-06-017).
Patient consent
The requirement for informed consent was waived due to the retrospective nature of the study.
Funding information
The authors received no financial support for the research, authorship, or publication of this article.