Research on Prediction of Preterm Birth Risk Based on Digital Twin Technology

Background: Preterm birth remains a major cause of perinatal morbidity and long-term developmental complications. Existing prediction methods often lack individualized assessment and have limited capability to integrate multi-source maternal-fetal information. This study aims to develop a personalized preterm birth risk prediction model and to construct a visual, interactive digital twin platform that enhances clinical communication and supports early risk identification. Methods: A total of 1157 structured clinical records collected from 2020 to 2024 were preprocessed through automated feature typing, missing-value handling, and normalization. Two complementary machine-learning models-FT-Transformer and Light Gradient Boosting Machine (LightGBM)-were trained and calibrated to produce probabilities. Their outputs were fused using a Stacking Logistic Regression framework to improve prediction stability and calibration. A 3D visualization module was developed using 3ds Max, PyQt6, and PyVista to generate personalized uterine-fetal models based on fetal position, placental location, and Biparietal Diameter (BPD), enabling synchronized display of prediction results. Results: The fused model achieved an AUC of 0.820, PR-AUC of 0.405, a Brier score of 0.040, and an expected calibration error (ECE) of 3.39 × 10^-3, demonstrating superior discrimination and probability reliability compared with single models. The interactive platform supports real-time data input, risk prediction, and adaptive 3D rendering, providing clear and intuitive visual feedback for clinical interpretation. Conclusions: The integration of machine learning fusion and digital twin visualization enables individualized assessment of preterm birth risk. The system improves model accuracy, enhances interpretability, and offers a practical tool for clinical follow-up, risk counseling, and maternal health education.