Microelements and biochemical biomarkers-based machine learning for predicting adverse pregnancy outcomes in Wilson's disease: risk stratification by integrating hepatic fibrosis and cerebral function.

BACKGROUND

Pregnancy in female patients with Wilson's disease (WD) raises significant gestational risks due to potential adverse pregnancy outcomes (APOs). This study developed machine learning (ML) algorithms based on microelement profiles and biochemical markers to identify APOs.

METHODS

Data on microelements (e.g., serum/urinary copper, iron), biochemical markers, and hepatic fibrosis were measured for all patients. Feature selection was performed using LASSO regression. Four ML models, including generalized linear model (GLM), deep learning (DL), random forest (RF), and gradient boosting machine (GBM), were developed and validated to distinguish between APOs and uneventful pregnancies (UP). Stratified analyses were conducted based on cerebral function (normal cerebral function vs. abnormal cerebral dysfunction) and hepatic fibrosis (with vs. without hepatic fibrosis).

RESULTS

114 patients with WD were enrolled, including 57 APO and 57 UP. The APO group exhibited a shorter disease duration, insufficient pre-pregnancy decoppering therapy, elevated levels of 24-h urinary copper and serum iron, and increased hepatic fibrosis biomarkers. Of the four ML models, the GLM had the highest accuracy (0.850) in the test set with excellent stability across training, test and validation sets, and no overfitting. RF and GBM had overfitting, while DL demonstrated poor generalization capability. Additionally, stratified analysis confirmed that the GLM showed strong robustness in most subgroups, whereas the GBM performed best performance in WD patients with cerebral dysfunction.

CONCLUSION

Microelements imbalance and hepatic fibrosis are associated with the risk of APOs in WD patients. The GLM, except for WD patients with cerebral dysfunction, serves as a reliable and generalizable predictive tool for APOs.