Recurrent spontaneous abortion (RSA), defined as ≥ 2 consecutive pregnancy losses before 24 weeks (ESHRE Guideline Group on on on et al., 2023), affects 1–5 % of reproductive-aged women (Dimitriadis et al., 2020, Quenby et al., 2021). Despite extensive evaluation including anatomical, genetic, thrombophilic, and endocrine assessments, up to 50 % of RSA cases remain unexplained (ESHRE Guideline Group on on on et al., 2023). Current guidelines lack validated predictive biomarkers, leading to delayed interventions. Emerging evidence suggests metabolic dysregulation and subclinical inflammation contribute to RSA pathogenesis (Sun et al., 2020). Women with RSA exhibit higher insulin resistance (Celik et al., 2011, Craig et al., 2002), altered β-cell function (Cao et al., 2025), significantly lower baseline testosterone levels (Zhao et al., 2025), and elevated neutrophil counts (Jiang et al., 2021). Recent research expands on inflammatory mechanisms in RSA. Studies indicate defective cytokine profiles, such as reduced IL-2 and IL-17A, in first-trimester miscarriage (Ku et al., 2023). Elevated fractalkine in amniotic fluid is associated with placental inflammation (Pala et al., 2024). However, maternal plasma cytokine screening shows limited predictive value for miscarriage, highlighting gaps in current biomarkers (Hannan et al., 2014).

While machine learning (ML) models have predicted obstetric complications, no model integrates metabolic, inflammatory, and clinical parameters for RSA prediction. Most existing models primarily focus on improving overall accuracy, but they lack the capacity to interpret the complex medical history and causative factors of individual cases. This limitation hampers the advancement of truly personalized treatment strategies.

We hypothesize that a multi-model machine learning approach can effectively predict RSA using routinely available clinical and laboratory data. The objective of this study is to identify key clinical and laboratory biomarkers that distinguish RSA patients from healthy controls. To this end, several machine learning models were developed and validated based on routinely collected clinical features. Model interpretability was achieved using the Shapley Additive Explanations(SHAP) framework, which allowed us to assess the contribution of individual variables to model predictions and to mimic clinical decision-making in RSA risk forecasting. Ultimately, this approach seeks to improve treatment strategies for RSA patients and mitigate the adverse outcomes associated with recurrent pregnancy management.