Developing and Validating a Machine Learning Approach for Prediction of Euploid Pregnancy Loss in the Copenhagen Pregnancy Loss Study

OBJECTIVE: To study whether a machine learning algorithm can effectively predict fetal genetic status, aneuploid or euploid, in cases of pregnancy loss, based solely on readily available clinical data. Accurate early prediction may enable improved clinical decision-making and personalized patient management.

DESIGN: Prospective multi-center cohort study within the Copenhagen Pregnancy Loss study, with a development cohort (n=788) from Copenhagen University Hospital Hvidovre and external validation cohorts from Copenhagen University Hospitals Herlev (n=229) and North Zealand (n=199). Enrollment was from November 2020 to May 2022.

SUBJECTS: Women ≥18 years with confirmed intrauterine pregnancy loss before 22 weeks at three Danish Copenhagen University Hospitals.

EXPOSURE: Ploidy status was determined using cell-free fetal DNA analysis from maternal blood. Forty-five clinical predictors including anthropometric data, medical history, lifestyle factors, and partner information were analyzed using a gradient boosted model. Feature importance was interpreted through Shapley additive explanation analysis to elucidate key prognostic indicators.

MAIN OUTCOME MEASURES: Model discrimination capability assessed by Area Under the Receiver Operating Characteristic curve and Area Under the Precision-Recall Curve, sensitivity, specificity, and calibration across cohorts.

RESULTS: The machine learning model demonstrated moderate discriminative ability with a cross-validated AUC-ROC of 0.69 (95% confidence interval 0.65-0.73) in the development cohort, maintaining equivalent performance in external validations. At a 90% specificity threshold, sensitivity was 54.4% in the development cohort (Hvidovre), 47% at Herlev, and 48% at North Zealand. Key predictors included maternal age, gestational age, paternal age and BMI, vitamin D supplementation, and vitamin E supplementation.

CONCLUSION: This study represents the first machine learning approach to distinguish euploid from aneuploid pregnancy losses using clinical data alone, providing a framework for earlier recognition of women with treatable conditions before multiple losses occur. Novel associations between vitamin supplementation and ploidy status invite further mechanistic investigation. However, moderate discriminative performance and population-specific factors highlight the need for additional validation and exploration before clinical integration.