Prediction of uterine dehiscence via machine learning by using lower uterine segment thickness and clinical features

BACKGROUND: With the global increase of cesarean delivery rates, the long-term effects of cesarean delivery have started to become clearer. One of the most prominent complications of cesarean delivery in recurrent pregnancies is uterine rupture. Assessing the risk of uterine rupture by accurately predicting dehiscence is very important to prevent untimely operations and/or maternal and fetal complications. OBJECTIVE: This study aimed to assess whether machine learning can be used to predict uterine dehiscence by using patients’ ultrasonographic findings, clinical findings, and demographic data as features. Hence, we investigated a potential method for preventing uterine rupture and its maternal and fetal complications. STUDY DESIGN: The study was conducted on 317 patients with term (>37 weeks’ gestation) singleton pregnancies and no medical complications or medications that may affect uterine wound healing. Demographics, body mass indices, smoking and drinking habits, clinical features, past pregnancies, number and history of abortions, interdelivery period, gestational week, number of previous cesarean deliveries, fetal presentation, fetal weight, tocography data, transabdominal ultrasonographic measurement of lower uterine segment full thickness and myometrium thickness, and lower uterine segment findings during cesarean delivery were collected and analyzed using machine learning techniques. Logistic regression, multilayer perceptron, support vector machine, random forest, and naive Bayes algorithms were used for classification. The dataset was evaluated using 10-fold cross-validation. Correct classification rate, F-score, Matthews correlation coefficient, precision-recall curve area, and receiver operating characteristic area were used as performance metrics. RESULTS: Among the machine learning techniques tested in this study, the naive Bayes algorithm showed the best predictive performance. Among the various combinations of features used for prediction, the essential features of parity, gravidity, tocographic contraction, cervical dilation, dilation and curettage, and sonographic thickness of lower uterine segment myometrium yielded the best results. The second-best performance was achieved with sonographic full thickness of lower uterine segment added to the base features. The base features alone could classify patients with 90.5% accuracy, whereas adding the myometrium measurement increased the classification performance by 5.1% to 95.6%. Adding the full thickness measurement to the base features raised the classification performance by 4.8% to 95.3% in terms of correct classification rate. CONCLUSION: The naive Bayes algorithm can correctly classify uterine dehiscence with a correct classification rate of 0.953, an F-score of 0.952, and a Matthews correlation coefficient value of 0.641. This result can be interpreted as indicating that by using clinical features and lower uterine segment ultrasonography findings, machine learning can be used to accurately predict uterine dehiscence.