Predicting a clinically narrow pelvis using neural network data analysis

Abstract

Aim: to improve the efficiency of predicting a clinically narrow pelvis (СNP) using neural network data analysis and to evaluate its prognostic characteristics.Materials and Мethods. The study was designed as a retrospective non-randomized clinical trial. An analysis of 184 born neonates was carried out: group 1 included 135 female patients whose delivery occurred through the natural birth canal, group 2 – 49 patients whose delivery was complicated by СNP development and ended up with emergency caesarean section. Examination of patients was carried out on the eve of childbirth (1–2 days) and included anamnesis, general and special obstetric examination, including pelvimetry, a clinical assessment of cephalopelvic disproportion was carried out during childbirth. The condition of newborns was assessed using the Apgar scale, height and body weight were measured. Neural network analysis was performed using the built-in Neural Networks module of SPSS Statistics Version 25.0 (IBM, USA).Results. Despite hypothetically important role of anatomically narrowed pelvis in development of cephalopelvic disproportion, no significant inter-group differences were found. Significant parameters (abdominal circumference, uterine fundus height and woman’s weight, fetal head circumference, as well as data on the presence or absence of oligohydramnios and fetal macrosomia) were determined, which were included in the test database to create the basis for training the multilayer perceptron. Out of 135 patients of group 1, the prognosis was negative in 131 (97.0 %), positive in 4 (3.0 %); out of 49 patients in group 2, negative in 0 (0.0 %), positive in 49 (100.0 %). The forecast accuracy of the developed model was 98 % (sensitivity – 100 %, specificity –97 %). The information content of neural network data analysis in СNP predicting is presented in ROC analysis: area under the curve (AUC) = 0.99 (95 % confidence interval = 0.97–1.00). Neonatal anthropometric parameters were significantly higher in group 2 vs. group 1, and the Apgar score at 1 minute was correspondingly lower.Conclusion. The use of neural network analysis of clinical data obtained on the eve of childbirth allows to predict СNP development at sufficient degree of accuracy (98.0 %), which, in the future, after being introduced into clinical practice, will optimize a choice of delivery method in patients at risk (anatomically narrow pelvis, large fetus), reduce emergency caesarean sections and improve birth outcomes.