# A multidimensional clinical prediction model for early screening of recurrent spontaneous abortion: integrating coagulation, immune, and endocrine markers

## TL;DR

This study develops a multidimensional model using coagulation, immune, and endocrine markers to accurately screen for recurrent spontaneous abortion risk in women.

## Contribution

The novel contribution is a Transformer-based model integrating six clinical domains for early RSA screening with high accuracy and cost-effectiveness.

## Key findings

- The TabPFN model achieved an AUC of 0.927 for RSA risk prediction using multidimensional clinical features.
- A six-biomarker model retained high discrimination (AUC = 0.925) with 83% accuracy using low-cost markers.
- Thrombo-immune dysregulation was identified as a central mechanism through SHAP analysis of key predictors.

## Abstract

Recurrent spontaneous abortion (RSA) affects 0.5%–2.5% of fertile couples and arises from complex, interacting thrombotic, immune, coagulation, endocrine–metabolic, and demographic factors. However, current early risk stratification in routine practice remains insufficient for population-level screening. We aimed to develop an accurate, low-cost, and clinically feasible early screening model for identifying women at high risk of RSA using routinely available clinical biomarkers.

This retrospective study enrolled women attending Guangdong Reproductive Hospital between 1 January 2020 and 31 December 2024. Among 1226 screened individuals, 285 met eligibility criteria and were included (181 RSA patients and 104 healthy controls). Demographic and laboratory variables were extracted from electronic medical records and structured follow-up. Ten classical machine-learning algorithms and a Transformer-based tabular model (TabPFN) were trained and compared. Class imbalance was handled using the synthetic minority oversampling technique (SMOTE). Model robustness was evaluated using 5-fold cross-validation. Biological-domain contributions were quantified through ablation analysis. Feature selection was optimized using recursive feature elimination with random forest (RFE-RF), and interpretability was assessed via SHAP.

The TabPFN Multidimensional model integrating features across six clinical domains achieved the best discriminative performance for RSA risk prediction (ROC–AUC = 0.927, 95% CI 0.891–0.947), outperforming all comparator algorithms. Domain ablation showed that removing any single biological category reduced performance, supporting the complementary value of multidimensional clinical integration. Acquired thrombophilia markers provided the strongest predictive contribution, followed by hereditary thrombophilia, immune indices, coagulation parameters, endocrine–metabolic variables, and demographic factors. A parsimonious six-biomarker model—anti-phosphatidylserine/prothrombin antibodies (aPS/PT), protein C (PC), antinuclear antibodies (ANA), antithrombin III (AT-III), thrombin time (TT), and body mass index (BMI)—retained high discrimination (AUC = 0.925) with 83% accuracy, supporting a pragmatic and cost-effective screening strategy. SHAP analysis identified elevated aPS/PT, ANA positivity, reduced AT-III activity, and prolonged TT as the most influential predictors, implicating thrombo-immune dysregulation as a central mechanism associated with RSA.

A Transformer-based tabular model using six routinely measured, low-cost biomarkers enable accurate, interpretable, and scalable early screening for RSA risk, with potential utility in resource-limited settings to facilitate timely referral and targeted preventive management.