Prediction of pregnancy outcomes in women with systemic lupus erythematosus before pregnancy: Application of machine learning models.

INTRODUCTION

Fetal loss is possible during pregnancy in women with systemic lupus erythematosus (SLE). Predicting pregnancy outcomes for women with SLE can be an effective aid in providing consultation and treatment services. Therefore, this study aimed to develop a machine-learning model that could predict pregnancy outcomes before pregnancy in women with SLE.

METHODS

The data of all pregnant women referred to the rheumatology center of Shariati Hospital and a specialized rheumatology clinic since 1980 were retrospectively collected from their medical records. Data collection was done by gathering 26 variables that affect pregnancy outcomes. Then, we used standard algorithms to select important features that affect pregnancy outcomes before pregnancy (11 different feature sets). A variety of machine learning algorithms were trained using both imbalanced and balanced datasets in Clementine and Weka software. Finally, the model with a higher area under the receiver operating characteristic curve (AUC) and F-score was selected to predict pregnancy outcomes.

RESULTS

Out of 149 pregnancies, 46 pregnancies resulted in spontaneous abortion, while 103 pregnancies resulted in live birth. Compared with other models, the Chi-square automatic interaction detection (CHAID) decision tree was selected as the best-performing model with higher accuracy (93.5 %), specificity (92.9 %), sensitivity (93.8 %), precision (97 %), F-score (0.95), and AUC (0.96).

CONCLUSION

By using the CHAID decision tree to predict the outcome of pregnancy in women with SLE and extracted rules, it is possible to use appropriate methods that prevent spontaneous abortion and also provide timely consultation to women with SLE for making decisions to become pregnant.