Collating Random Forest Classifier and Artificial Neural Networks for the Risk Detection of Maternal Health

This paper compares the performance of ANN models to Random Forests on benchmarking a test dataset against the risk levels of maternal health as low, mid, or high. The features considered are clinical features, namely age, blood pressure, and heart rate. The dataset was imbalanced and needed adapting the ANN architecture by fitting class weights and dropout regularization. The best test accuracy obtained in this fashion was 73%. As the ANN had learned many of the complex data patterns in the dataset, it was, however, still limited by the moderate size and skewness of the dataset. The Random Forest model with grid search produced a much better accuracy of 85.71% compared with ANN, thereby suggesting relatively high generalization across all the classes for better minority risk level classes. An ensemble approach of the Random Forest method appears to better serve toward high accuracy along with interpretability for clinical decision support in maternal health risk assessment.