Background: Pregnancy-associated hypertension (HTN) is one of the most serious complications during pregnancy. Prophylactic aspirin medication in early pregnancy has been reported to reduce effectively the risk of pregnancy-associated HTN for high-risk women. A few studies tried to predict pregnancy-associated HTN using machine learning methods but failed to show robust performance. Here, we selected important clinical variables using feature selection methods and developed a prediction model for pregnancy-associated HTN using graph-based semi-supervised learning (SSL). 
Methods: This is a secondary analysis from a prospective study of healthy pregnant women. For ranking/selecting variables, four feature selection methods were applied to elucidate between 32 routine clinical variables in the first trimester of pregnancy and pregnancy-associated HTN. Then, a patient network with selected variables was constructed to recognize patients with similar patterns. To predict pregnancy-associated HTN with exploiting the underlying network structure of patients, a graph-based SSL was employed.
Results: A patient network was constructed with 1,401 pregnant women, of which 33 women had developed pregnancy-associated HTN. 11 clinical factors (out of 32 routine variables) were selected from the feature selections. We compared the prediction performances with four machine learning methods, including graph-based SSL, logistic regression, support vector machine, and random forest, with 5-fold cross validations. The graph-based SSL using selected 11 clinical factors has achieved the best prediction performance with an average AUROC of 0.89, compared to logistic regression (avg. AUROC of 0.84). 
Conclusion: Graph-based SSL with 11 selected variables can produce a more accurate prediction result for pregnancy-associated HTN in early pregnancy.