Machine learning for predictive risk stratification in recurrent miscarriage: a systematic review

Abstract Background Recurrent miscarriage (RM), affecting 1–2% of couples, often lacks a clear etiology, complicating prognosis and care. Machine learning (ML) offers a potential paradigm shift by enabling robust risk stratification. This systematic review synthesizes the current landscape of ML applications for predicting RM, evaluating methodological quality, performance, and clinical translatability. Methods We systematically searched PubMed/MEDLINE, Scopus, and Web of Science until 9 December 2024. Studies developing or validating ML models for predicting RM or related pregnancy loss were included. Data on study characteristics, datasets, predictive features, ML methodologies, validation approaches, and performance metrics were extracted. Quality was assessed using the CHARMS tool. Results Twenty-six studies were included. Most were retrospective (n = 24) and originated from China (n = 14). Dataset sizes varied widely (n ≈ 78 to 338,904). Models incorporated diverse data, from clinical demographics (85% of studies) to embryological morphokinetics and omics biomarkers (26%). Tree-based ensemble algorithms, particularly Random Forest and XGBoost, demonstrated consistently high performance. While reported discriminative performance was often promising (AUC frequently > 0.80), critical methodological gaps were identified. There was a pervasive lack of robust external validation; most studies relied solely on internal validation. Reporting of calibration and handling of missing data was frequently inadequate, hindering assessment of clinical reliability. Conclusion ML shows substantial promise for transforming RM prediction through multi-factorial risk stratification. However, the field’s current evolution from proof-of-concept to clinical tool is hampered by a significant validation and reproducibility gap. Future work must prioritize external validation, standardized reporting per TRIPOD-ML guidelines, and demonstration of clinical utility to bridge this gap and realize the potential of ML for improving patient care. Prospero registration CRD420251176304. Clinical trial number Not applicable.