Routine prenatal screening through fetal ultrasound forms the backbone of modern obstetric practice, offering clinicians a non-invasive window into early developmental abnormalities and chromosomal disorders. Yet the accuracy of these scans depends heavily on the availability of skilled sonographers, and this reliance inevitably introduces variability in diagnostic outcomes, particularly in regions where specialist access is limited. To address these challenges, we developed a hybrid AI framework that couples a fine-tuned EfficientNetB0 convolutional neural network (CNN) with an XGBoost ensemble classifier for comprehensive fetal health screening. In our two-stage pipeline, the CNN first performs binary classification on uploaded ultrasound images, flagging potential developmental anomalies. When an abnormality is detected, a trained XGBoost model then works on five maternal clinical features—age, diabetes status, hypertension, family history of chromosomal disorders, and gestational week—to differentiate among Down Syndrome (Trisomy 21), Edward Syndrome (Trisomy 18), Patau Syndrome (Trisomy 13), and Turner Syndrome (Monosomy X). We deployed the complete system as an interactive Streamlit web dashboard, which outputs real-time risk scores, condition-wise probabilities, and evidence-based clinical guidance. Evaluations on a curated ultrasound image dataset together with synthetic maternal records yielded a CNN accuracy of 97.2%, precision of 96.4%, recall of 95.8%, and F1-score of 96.1%. The XGBoost subtype classifier achieved 94.6% multi-class accuracy. Together, these figures suggest the proposed architecture is both clinically viable and accessible—particularly suited to support practitioners in resource-constrained environments.

Keywords: Fetal health screening; deep learning; EfficientNetB0; XGBoost; ultrasound image classification; chromosomal abnormalities; prenatal diagnosis; convolutional neural network; hybrid model; Streamlit deployment.