Predicting adverse perinatal outcomes in small-for-gestational-age fetuses using MRI, ultrasound and clinical data

Abstract Background Fetal growth restriction (FGR) is associated with adverse perinatal outcomes. Existing sonographic approaches offer limited predictive accuracy. Combining fetal MRI, ultrasound and clinical data may improve perinatal prognostication. Purpose To evaluate whether integrating prenatal MRI, ultrasound, and clinical features using machine learning (ML) improves prediction of adverse perinatal outcomes in FGR or small for gestational age (SGA) pregnancies. Materials and Methods This single-center study included prospectively enrolled FGR/SGA and retrospectively included appropriate-for-gestational-age cases, with follow-up through neonatal discharge. Twenty-seven features from MRI, ultrasound and clinical data were used in the final analysis. Seven ML classifiers were trained using stratified 5-fold cross-validation to predict composite adverse neonatal outcomes (CANO) and non-reassuring fetal status (NRFS). Sensitivity and specificity of the top-performing model (based on area under the curve AUC) were compared to standard biometric thresholds (estimated fetal weight and/or abdominal circumference 10th/3rd centiles). Multiparametric, MRI-only and ultrasound-only models were compared, along with reduced models using four features for CANO and two for NRFS. Results 131 participants were included (60 FGR/SGA, 71 appropriate-for-gestational-age). The random forest method achieved the highest AUC for predicting CANO (0.912; 95% confidence interval CI, 0.83-0.99) and NRFS (0.834; 95% CI, 0.76-0.91). For CANO, the multiparametric model demonstrated a 25% higher sensitivity (P = 0.005) and 17% higher specificity (P 0.001) compared with the 3rd centile threshold, and improved specificity over the 10th centile threshold by 29% (P 0.001). Sensitivity did not differ significantly from the 10th centile threshold (P = 0.366). For NRFS, specificity increased by 26% and 40% over the 3rd and 10th centile thresholds, respectively (P 0.001), without significant differences in sensitivity (P = 1). No statistically significant differences were observed between the multiparametric, ultrasound-only, and MRI-only models (P ≥ 0.826), or between full and reduced models (P ≥ 0.313). Conclusions ML-based models integrating multimodal data may improve risk stratification for predicting adverse perinatal outcomes in FGR/SGA pregnancies.