Prediction of Preeclampsia and Intrauterine Growth Restriction: Development of Machine Learning Models on a Prospective Cohort

• DOI: 10.2196/15411
• Publication Date: 2020-05-04T16:06:17Z
• Authors: Herdiantri Sufriyana; Yu-Wei Wu; Emily Chia-Yu Su
• Keywords: Predictive modelling; Intrauterine growth restriction

Background Preeclampsia and intrauterine growth restriction are placental dysfunction–related disorders (PDDs) that require a referral decision be made within certain time period. An appropriate prediction model should developed for these diseases. However, previous models did not demonstrate robust performances and/or they were from datasets with highly imbalanced classes. Objective In this study, we predictive of PDDs by machine learning uses features at 24-37 weeks’ gestation, including maternal characteristics, uterine artery (UtA) Doppler measures, soluble fms-like tyrosine kinase receptor-1 (sFlt-1), factor (PlGF). Methods A public dataset was taken prospective cohort study included pregnant women (66/95, 69%) control group (29/95, 31%). Preliminary selection based on statistical analysis using SAS 9.4 (SAS Institute). We used Weka (Waikato Environment Knowledge Analysis) 3.8.3 (The University Waikato, Hamilton, NZ) to automatically select the best its optimization algorithm. also manually selected 23 white-box models. Models, those recent studies, compared interval estimation evaluation metrics. Matthew correlation coefficient (MCC) as main metric. It is overoptimistic evaluate performance class imbalance. Repeated 10-fold cross-validation applied. Results The classification via regression chosen model. Our had MCC (.93, 95% CI .87-1.00, vs .64, .57-.71) specificity (100%, 100-100, 90%, 90-90) each metric studies. sensitivity inferior (95%, 91-100, 100%, 92-100). area under receiver operating characteristic curve competitive (0.970, 0.966-0.974, 0.987, 0.980-0.994). Features in weight, BMI, pulsatility index UtA, sFlt-1, PlGF. most important feature sFlt-1/PlGF ratio. This an M5P algorithm consisting tree four linear different thresholds. better than ones among studies terms balance size case 69%, 27/239, 11.3%). Conclusions performance. deal problem context clinical management, may improve mortality neonatal morbidity reduce health care costs.