Prediction of Shoulder Dystocia Using Maternal Pelvic and Fetal MRI with Machine Learning

-Background Shoulder dystocia (SD) is a rare but severe complication of vaginal birth, associated with substantial neonatal and maternal morbidity. Existing clinical and ultrasonographic predictors show limited discriminatory accuracy. Magnetic resonance imaging (MRI) offers detailed assessment of fetal and maternal pelvic morphology, but its contribution to SD prediction has not been systematically evaluated. We aimed to develop and validate machine-learning models incorporating maternal and fetal MRI features to improve prediction of SD. -Methods We conducted a retrospective diagnostic-prediction study including singleton pregnancies between 35+0 and 37+0 weeks of gestation undergoing MRI at a tertiary centre between Jan 1, 2015, and Dec 31, 2024. Maternal demographics, obstetric history, ultrasound fetal biometry, and MRI-derived fetal and pelvic morphometrics were collected. SD was defined according to Royal College of Obstetricians and Gynaecologists criteria. To address the competing risk of intrapartum cesarean delivery (CD), prediction used a sequential two-stage modelling framework: stage 1 estimated the probability of CD, and stage 2 estimated the conditional probability of SD among vaginal births. Three prespecified predictor groups were evaluated: (1) clinical variables and ultrasound biometry, (2) clinical variables plus estimated fetal weight percentile, and (3) clinical variables plus MRI morphometrics. Models were developed using 20 multiply imputed training datasets and internally validated by cross-validation; external performance was assessed in a held-out test set. Discrimination was summarized using area under the receiver operating characteristic curve (ROC-AUC) and area under the precision-recall curve (AUCPR). -Findings Among 2195 pregnancies analyzed, 1929 (88.0%) had no SD or intrapartum CD, 203 (9.2%) had intrapartum CD, and 63 (2.9%) had SD. Clinical and ultrasound-based models showed moderate predictive ability (ROC-AUC 0.79 and AUCPR 0.14 for SD). MRI-enhanced models substantially improved SD prediction (ROC-AUC 0.85; AUCPR 0.38 in cross-validation) and retained high accuracy in the independent test set (ROC-AUC 0.93; AUCPR 0.35). Gains were consistent across strata of fetal size, maternal origin, and obstetric history. -Interpretation Machine-learning models integrating maternal pelvic and fetal morphometry from late-pregnancy MRI markedly improved prediction of SD compared with clinical or ultrasound-based approaches. These findings support a potential role for advanced imaging and computational methods in antenatal risk stratification, although multicentre validation is needed before clinical implementation.