Abstract
<title>Abstract</title> <p>Objective Many adverse pregnancy outcomes (APOs) arise from placental dysfunction that begins before clinical disease is apparent. We developed PUMA-Net (Placental Ultrasound Multi-scale Attention Network), a deep learning model designed to extract subtle pathological information from routine placental ultrasound images, and examined whether combining image-derived risk with clinical variables could improve early identification of APOs. Materials and Methods We included 647 pregnant women who received regular antenatal care at two centers and had first-trimester and second-trimester placental ultrasound images available, together with clinical data. At Center 1, 407 retrospectively enrolled women formed the training cohort and 133 prospectively enrolled women formed the internal validation cohort. Center 2 provided an independent external validation cohort of 107 retrospectively enrolled women. PUMA-Net used a convolutional backbone with attention modules and a multiscale feature pyramid to capture focal and heterogeneous placental features. Prediction probabilities were generated separately from first- and second-trimester images. Candidate clinical predictors were assessed by multivariable logistic regression, and the final combined model incorporated PUMA-Net probabilities from the first and second trimesters and the selected clinical predictor in a logistic-regression fusion model. We evaluated model performance using receiver operating characteristic analysis, calibration plots, and decision curve analysis. Results Prepregnancy BMI was the only independent clinical predictor in multivariable analysis. The model that combined PUMA-Net prediction probabilities with prepregnancy BMI performed best. Its AUC was 0.899 in the training cohort and 0.875 in the internal validation cohort. In the external validation cohort, the combined model retained good discrimination, with an AUC of 0.836. Calibration plots showed close agreement between predicted and observed risk. On DCA, the combined model offered greater net benefit across clinically relevant threshold probabilities than either treating all patients as high risk or treating all as low risk. Conclusions PUMA-Net captured placental ultrasound features that were associated with subsequent APOs, and its combination with prepregnancy BMI improved risk prediction across internal and external validation cohorts. This multimodal approach may help clinicians identify pregnancies that warrant closer surveillance while reassuring women at lower predicted risk.</p>