Machine learning-based integration of antiphospholipid antibodies for predicting pregnancy outcomes in patients with recurrent spontaneous abortion: a multicenter prospective study.

OBJECTIVES

Antiphospholipid antibodies (aPLs) are closely associated with recurrent spontaneous abortion (RSA) and adverse pregnancy outcomes (POs). Non-criteria aPLs may aid in diagnosing seronegative aPL-related RSA, but their limited performance often leads to missed or delayed diagnosis. This study aimed to improve diagnostic and predictive accuracy by integrating criteria and non-criteria aPLs with machine learning (ML) algorithm.

METHODS

In this multicenter prospective study, 1,321 participants were recruited and 751 included. Fifteen aPLs were measured using chemiluminescence immunoassay. Six ML algorithms were trained, and the optimal model was evaluated using the area under the curve (AUC), calibration curves, decision curve analysis (DCA), and Shapley additive explanations (SHAP).

RESULTS

Using the 95th percentile as the cut-off, aPL positivity ranged from 2.62 to 12.13 % in pregnant woman with a history of RSA and 4.29-19.74 % in non-pregnant woman with a history of RSA. The light gradient boosting machine (LGBM) model achieved AUCs of 0.875 (pregnant) and 0.778 (non-pregnant) for RSA prediction, while the random forest (RF) model achieved an AUC of 0.885 for PO prediction - surpassing all single indicators (AUC 0.516-0.647). Calibration, DCA, and SHAP analyses demonstrated strong clinical utility.

CONCLUSIONS

The ML models substantially improved diagnostic and predictive performance for RSA and PO. The LGBM and RF models showed the best accuracy and may serve as auxiliary diagnostic and early warning tools. Larger external cohorts are needed for further validation.