Giving more insight for automatic risk prediction during pregnancy with interpretable machine learning

Abstract

Maternal mortality rate (MMR) in Indonesia intercensal population survey (SUPAS) was considered high. For pregnancy risk detection, the public health center (puskesmas) applies a Poedji Rochjati screening card (KSPR) demonstrating 20 features. In addition to KSPR, pregnancy risk monitoring has been assisted with a pregnancy control card. Because of the differences in the number of features between the two control cards, it is necessary to make agreements between them. Our objectives are determining the most influential features, exploring the links among features on the KSPR and pregnancy control cards, and building a machine learning model for predicting pregnancy risk. For the first objective, we use correlation-based feature selection (CFS) and C5.0 algorithm. The next objective was answered by the union operation in the features produced by the two techniques. By performing the machine learning experiment on these features, the accuracy of the XGBoost algorithm demonstrated the hightest results of 94% followed by random forest, Naïve Bayes, and k-Nearest neighbor algorithms, 87%, 66%, and 60% respectively. Interpretability aspects are implemented with SHAP and LIME to provide more insight for classification model. In conclusion, the similarity feature generated in the two interpretation approaches confirmed that Cesar was dominant in determining pregnancy risk.