Application of Machine Learning and Cervical Impedance Analyses to Preterm Birth Prediction

Objectives: Preterm birth (PTB) is the second leading cause of premature child deaths under 5 years old worldwide and is associated with a series of lifetime disease and disability and substantial long-term healthcare costs. Prediction and prevention of PTB remains limited by the recognition of PTB causes and the modest accuracy of current prediction approaches. This work was based on the Electrical Impedance Prediction of Preterm birth by spectroscopy of the cervix (ECCLIPPxTM) project and the Electrical Impedance Prediction of Preterm Birth by Spectroscopy of the cervix II (ECCLIPPx II) project, which proposed the novel PTB prediction approaches using cervical electrical impedance spectroscopy (EIS) and cervical magnetic impedance spectroscopy (MIS) data. The aim was to apply machine learning and cervical impedance analysis to preterm birth prediction. The objectives include (1) pre-process EIS data and carry out the EIS based PTB prediction, (2) improve the PTB prediction accuracy by combining EIS-based and maternal characteristics-based PTB risk models and (3) carry out a feasibility study on the MIS-based PTB risk prediction. Methods: For the study of EIS data based PTB prediction, the EIS data from 438 recruited women at 19 to 23 weeks of gestation were investigated. A principle of identifying the optimum EIS spectrum among several recorded EIS spectra for each patient was proposed preparing for training the optimal-EIS spectrum filter model. Then the association of the EIS data with the PTB risk was determined by training the logistic regression (LR) model using a series of data modelling and analysis techniques. Moreover, a models-combining approach of combining two probabilities given by EIS-based PTB risk prediction model and a maternal characteristics-based PTB risk prediction model, was proposed to improve accuracy of prediction. For the study on MIS data based PTB prediction, the MIS data from 84 recruited women at 18 to 35 weeks of gestation were used. Two calibration methods were proposed to pre-process the MIS data. Then, the support vector machine (SVM) classifier was applied to determine the association of MIS data with the PTB risk. Conclusions: The idea of building optimal-EIS spectrum filter model helped process the heterogeneous EIS records. After that, the EIS-based PTB risk prediction model and the combined model by an EIS-based PTB risk prediction model and a maternal characteristics-based PTB risk prediction model had excellent predicting abilities for both high-risk and low-risk women, within one to 23 weeks. The proposed MIS data calibration methods can reduce the MIS device’s effects on MIS data, by removing components unrelated to cervix. MIS-based PTB risk prediction models had good training performances, so that MIS-based PTB risk prediction is worth a further study.