A filter-predictor polynomial feature based machine learning approach to predicting preterm birth from cervical electrical impedance spectroscopy

Abstract

Objective:Preterm birth (PTB) refers to the delivery of a baby before 37 weeks of pregnancy. PTB can cause new-born deaths and long term diseases of children. The objective of this work is to propose a novel machine learning approach to predict PTB from cervical electrical impedance spectroscopy (EIS) of pregnant women during 20 to 22 weeks of pregnancy. Methods:The proposed approach selects a best EIS spectrum using a filter, then, predicts PTB based on the selected EIS spectrum using a predictor. A methodology is proposed to train a polynomial feature based logistic regression (PFLR) and a polynomial feature based random forest (PFRF) as filters or predictors, respectively, from an imbalanced high dimensional EIS dataset. A rough set-based genetic algorithm for selecting optimal polynomial feature subsets, a pre-processing step in the proposed methodology, is proposed. Results:For an EIS dataset of 438 patients of various demographics, previous obstetric and treatment history, PFRF achieves an average test set AUC 0.76 and PFLR achieves an average test set AUC 0.74. For the 365 patients who received no treatment interventions during pregnancy, PFRF achieves an average test set AUC 0.8 and PFLR achieves an average test set AUC 0.79. The proposed approach has outperformed the existing PTB prediction approaches based on EIS, demographics, previous obstetric history, TVUS CL, and FFN and RBF kernel SVM, MLPs (1-hidden-layer, 2-hidden-layers and 3-hidden-layers), XGBoost, logistic regression and random forest. Conclusion and significance:The proposed approach has demonstrated its potential utility in clinical practice.