This retrospective study included patients who were diagnosed with cervical insufficiency during their first pregnancy and had a second delivery at Xiamen Humanity Maternity Hospital between January 2019 and January 2024. The inclusion criteria were as follows: age $\ge$18 years; diagnosed with cervical insufficiency during the first pregnancy; completion of a subsequent second pregnancy and delivery; and availability of complete clinical data. The exclusion criteria were as follows: multiple pregnancies; severe pregnancy complications that could significantly affect pregnancy outcomes; presence of severe systemic diseases; major fetal malformations or chromosomal abnormalities; severe hepatic or renal dysfunction; severe mental illness or poor compliance; and missing critical data.

For samples with a small amount of missing data, regression imputation (or multiple imputation) based on other variables was performed. Samples with a substantial amount of missing data, such as those with more than 30% and not missing completely at random, were excluded from the analysis.

According to the inclusion criteria, a total of 550 pregnant women were initially deemed eligible for the study. After applying the exclusion criteria, a total of 480 women were ultimately included in the study. Among them, 86 experienced preterm birth, and 110 were diagnosed with cervical insufficiency during their second pregnancy. All participants were categorized into a preterm birth group and a non-preterm birth group based on the occurrence of preterm birth during their second pregnancy.

Patient data collection included maternal physiological and health status (age, body mass index [BMI], history of uterine cavity surgery, history of cervical conization surgery, history of uterine surgery, and use of folate and iron supplementation); relevant information from the first pregnancy (whether cervical cerclage was performed and its type, premature rupture of membranes [PROM], pregnancy-induced hypertension, gestational diabetes, placental abruption, placenta previa, mode of delivery, occurrence of preterm birth, cervical laceration, and the degree of cervical softening or dilation); measurement of cervical length at different gestational weeks during the first pregnancy (16–18 weeks, 20–22 weeks, and 22–24 weeks); the interpregnancy interval; and fetal related characteristics (including birth weight, fetal position, and amniotic fluid index).

Cervical insufficiency: According to the Clinical Practice Guideline No. 373 on Cervical Insufficiency and Cerclage, published by the Society of Obstetricians and Gynaecologists of Canada (SOGC) in 2019 [12], cervical insufficiency is defined as a cervical length of less than 25 mm during the mid-trimester in the absence of clinical signs of preterm labor. It is typically characterized by painless, progressive cervical dilation in the second or third trimester, with or without PROM or protrusion of the amniotic sac through the cervix.

Lasso regression was used to screen significant variables. The optimal regularization parameter (lambda) was selected through cross-validation. The optimal lambda value was 0.0047 for the preterm birth model and 0.0087 for the cervical insufficiency model.

The dataset was randomly split into a training set and a test set at a ratio of 7:3. Within the training set, 10-fold cross-validation was used to optimize the parameters of the LR model. For the XGBoost model, hyperparameters were tuned using grid search, which systematically explores combinations of hyperparameters (e.g., learning rate, maximum tree depth, and subsample ratio) to identify the optimal configuration. The final LR and XGBoost models were constructed using the optimal parameter settings. Model performance was evaluated on the test set using area under the curve (AUC), precision, accuracy, recall, and F1 score.

Key predictive features with high importance in both the LR and XGBoost models were selected to construct the risk model. LR coefficients and SHAP values from the XGBoost model were standardized, and the overall risk score was calculated as:

Risk Model = $\sum$(Factor_i $\times$ Coef_i)

where Factor represents the value of each selected feature, and Coef represents the standardized weight.

The predictive performance of the risk model was assessed using receiver operating characteristic (ROC) curves, with evaluation metrics such as AUC, sensitivity, specificity, optimal threshold, and the Youden Index. Additionally, the model’s applicability to predicting the risk of both preterm birth and cervical insufficiency was evaluated.

All analyses in this study were performed using R software version 4.4.0 (R Foundation for Statistical, Vienna, Austria). Continuous data were presented as mean $\pm$ SD or median (minimum–maximum) and compared using either the t-test or Mann-Whitney U test, depending on data distribution. Categorical data were expressed as frequency (percentage). For the analysis of categorical variables, Fisher’s exact test was used when the theoretical frequency (T) was less than 1. When 1 $\le$ T 5 and the total sample size was $\ge$40, the continuity-corrected chi-square test was applied.

Compared with pregnant women who did not experience preterm birth in their second pregnancy, pregnant those who experienced preterm birth exhibited statistically significant differences in several baseline characteristics. The incidence of a history of uterine cavity surgery was higher in the preterm birth group (13.95% vs. 6.85%, p = 0.049). The proportion of emergency cervical cerclage was significantly higher in the preterm birth group (6.98% vs. 2.03%, p = 0.017). The incidence of PROM was higher in the preterm group (20.93% vs. 10.66%, p = 0.015). The proportion of moderate and severe cervical lacerations was significantly higher in the preterm group compared to the non-preterm group (moderate: 25.58% vs. 17.77%; severe: 17.44% vs. 7.87%, p = 0.002). The proportion of moderate cervical softening or dilation abnormalities was significantly higher in the preterm birth group compared to the non-preterm group (22.09% vs. 8.88%, p = 0.002). Additionally, cervical length at both 16–18 weeks and 22–24 weeks of gestation was significantly shorter in the preterm group (16–18 weeks: 27.70 vs. 29.40 mm, p 0.001; 22–24 weeks: 25.00 vs. 26.20 mm, p = 0.004). In addition, the preterm birth group had a higher proportion of preterm birth history during the first pregnancy (36.05% vs. 18.78%, p = 0.001), and the incidence of abnormal fetal position (breech presentation) was significantly increased (4.65% vs. 1.02%, p = 0.045) (Table 1). The median gestational age at intervention was 13 weeks (range: 12–14 weeks) in the prophylactic cerclage group and 20 weeks (range: 16–25 weeks) in the emergency cerclage group. The usage rate of Mersilene tape was 67.94%, while Prolene suture was used in 32.06% of cases. The McDonald suturing technique accounted for 88.55% of procedures, whereas the Shirodkar technique was used in 11.45%. In most cases, cervical cerclage was performed independently by an experienced obstetrician. In more complex cases, two operators were involved. No cervical trauma was observed during any of the procedures in this study.

The optimal lambda value for the Lasso model predicting preterm birth was 0.004735971, while the optimal lambda value for the Lasso model predicting cervical insufficiency was 0.008738651. The features selected for the Lasso model predicting preterm birth included: prophylactic cerclage, emergency cerclage, PROM, preterm birth, cervical laceration, failure of the cervix to soften or dilate normally, cervical length at 16 to 18 weeks, and cervical length at 22 to 24 weeks. The selected features for the Lasso model predicting cervical insufficiency included: history of uterine cavity surgery, prophylactic cerclage, emergency cerclage, PROM, preterm birth, cervical laceration, failure of the cervix to soften or dilate normally, fetal presentation (breech presentation), cervical length at 22 to 24 weeks, and cervical length at 16 to 18 weeks

The performance of the four models varied. For preterm birth prediction, the XGBoost model achieved an AUC of 0.745, slightly higher than the LR model’s AUC of 0.729. Additionally, XGBoost outperformed LR in precision, accuracy, and F1 score, demonstrating strong predictive ability (Fig. 1A,B). In comparison, the LR model had a precision of 0.706 and an F1 score of 0.749. Although the LR model performed well overall, it was still inferior to XGBoost. For the prediction of cervical insufficiency, the LR model achieved and AUC of 0.777, demonstrating high specificity (0.885) but a relatively low recall rate (0.595), resulting in an F1 score of 0.695. The overall performance was acceptable but somewhat limited. The XGBoost model achieved an AUC of 0.711, with a moderate recall rate (0.645) but low specificity (0.553) (Fig. 1C,D). XGBoost performs slightly worse than LR in predicting cervical insufficiency; however, its higher recall rate provides a greater number of positive samples for identifying preterm birth (Table 2). Additionally, we found that although breech presentation was identified as an independent risk factor for predicting cervical insufficiency, its odds ratio (OR) was exceptionally high and accompanied by a wide confidence interval (CI). This may be due to the relatively small number of breech presentations in our dataset, leading to an unstable estimation of this variable within the model. Therefore, caution is warranted when interpreting or using this variable for predictive purposes.

Fig. 1.

ROC curves for predicting preterm birth and cervical insufficiency using different models. (A) LR model predicting preterm birth. (B) XGBoost model predicting preterm birth. (C) LR model predicting cervical insufficiency. (D) XGBoost model predicting cervical insufficiency. AUC, area under the curve; ROC, receiver operating characteristic; LR, logistic regression; XGBoost, eXtreme Gradient Boosting.

The results of LR analysis indicate that prophylactic cervical cerclage is associated with a reduced of preterm birth (OR = 0.619, p = 0.042), suggesting that patients who undergo this procedure during their first pregnancy have a lower risk of preterm birth during their second pregnancy. Conversely, emergency cervical cerclage is significantly positively correlated with the risk of preterm birth (OR = 9.211, p = 0.001), indicating that patients who undergo this procedure during their first pregnancy have a higher likelihood of preterm birth during their second pregnancy. PROM also significantly increases the risk of preterm birth (OR = 1.946, p = 0.015). In addition, cervical laceration and failure to soften or dilate normally were also significantly associated with an increased risk of preterm birth (OR = 1.714, p 0.001 and OR = 2.215, p 0.001). Cervical length at 16–18 weeks and 22–24 weeks was significantly negatively correlated with preterm birth (OR = 0.434, p = 0.004 and OR = 0.431, p 0.001, respectively), indicating that a shorter cervical length during the first pregnancy may increase the risk of preterm birth in the second pregnancy.

For predictive factors of cervical insufficiency, a history of uterine cavity surgery was significantly associated with an increased risk (OR = 3.842, p = 0.009). PROM also significantly increased the risk of cervical insufficiency (OR = 5.290, p 0.001), as did a history of preterm birth (OR = 4.431, p 0.001). Cervical laceration and failure to soften or dilate normally also had significant effects on cervical insufficiency (OR = 1.996, p = 0.002 and OR = 2.354, p = 0.011). In addition, the breech presentation significantly increased the risk of cervical insufficiency (OR = 29.285, p = 0.011, respectively). The cervical length at 16–18 weeks and 22–24 weeks was negatively associated with cervical insufficiency, especially at 22–24 weeks (OR = 0.113, p = 0.023 and OR = 0.040, p = 0.008, respectively), indicating that a shorter cervical length during the first pregnancy is closely related to the occurrence of cervical insufficiency in the second pregnancy.

In summary, the occurrence of preterm birth and cervical insufficiency in the second pregnancy is influenced by cervical condition, cervical surgery, and pregnancy-related factors during the first pregnancy, with changes in cervical length playing an important role in both outcomes (Table 3).

The XGBoost model was interpreted using SHAP visualization analysis, which revealed that for preterm birth, key predictive features include cervical length at 16–18 weeks and 22–24 weeks during the first pregnancy, cervical softening or dilation at the first delivery, cervical laceration, and a history of preterm birth in the first pregnancy. These factors significantly influence the risk of preterm birth in the subsequent pregnancy. For cervical insufficiency, key predictive factors included length of the cervix at 16–18 weeks and 22–24 weeks during the first pregnancy, history of preterm birth during the first pregnancy, cervical lacerations, and PROM. These features significantly influence the risk of cervical insufficiency in the subsequent pregnancy (Fig. 2A,B).

Fig. 2.

SHAP visualization analysis of XGBoost models for pregnancy outcomes. (A) Prediction of preterm birth. (B) Prediction of cervical insufficiency. SHAP, SHapley Additive exPlanations.

Four characteristic factors were selected to construct the risk model, namely cervical length at 16–18 weeks of the first pregnancy, cervical length at 22–24 weeks, history of preterm birth, and cervical laceration. Subsequently, the ROC curve was used to evaluate the performance of this model in predicting preterm birth and cervical insufficiency. The results showed that the risk model for preterm birth achieved an AUC of 0.711, indicating good predictive performance, with a 95% CI of 0.653 to 0.768. The optimal threshold is 40.285, Youden index is 0.407, sensitivity is 0.779, and specificity is 0.628, indicating that the model has a good balance of sensitivity and specificity in predicting premature birth. For the risk model of cervical incompetence, the AUC value is 0.686, indicating that the predictive ability of the model is moderate, with a 95% CI of 0.628 to 0.745. The optimal threshold is 41.039, with a Youden index is 0.396, sensitivity of 0.733, and specificity of 0.663. These results indicate that the model has high sensitivity but relatively lower specificity in predicting cervical insufficiency, suggesting a potential risk of false positives (Table 2) (Fig. 3A,B).

Fig. 3.

ROC curves for prediction models. (A) ROC curve of risk model predicting preterm birth. (B) ROC curve for predicting cervical incompetence using a risk model. ROC, receiver operating characteristic.