Multivariate adaptive regression splines analysis to predict biomarkers of spontaneous preterm birth

摘要

Objective To develop classification models of demographic/clinical factors and biomarker data from spontaneous preterm birth in African Americans and Caucasians. Design Secondary analysis of biomarker data using multivariate adaptive regression splines (MARS), a supervised machine learning algorithm method. Setting Analysis of data on 36 biomarkers from 191 women was reduced by MARS to develop predictive models for preterm birth in African Americans and Caucasians. Samples Maternal plasma, cord plasma collected at admission for preterm or term labor and amniotic fluid at delivery. Methods Data were partitioned into training and testing sets. Variable importance, a relative indicator (0-100%) and area under the receiver operating characteristic curve (AUC) characterized results. Results Multivariate adaptive regression splines generated models for combined and racially stratified biomarker data. Clinical and demographic data did not contribute to the model. Racial stratification of data produced distinct models in all three compartments. In African Americans maternal plasma samples IL-1RA, TNF-α, angiopoietin 2, TNFRI, IL-5, MIP1α, IL-1β and TGF-α modeled preterm birth (AUC train: 0.98, AUC test: 0.86). In Caucasians TNFR1, ICAM-1 and IL-1RA contributed to the model (AUC train: 0.84, AUC test: 0.68). African Americans cord plasma samples produced IL-12P70, IL-8 (AUC train: 0.82, AUC test: 0.66). Cord plasma in Caucasians modeled IGFII, PDGFBB, TGF-β1, IL-12P70, and TIMP1 (AUC train: 0.99, AUC test: 0.82). Amniotic fluid in African Americans modeled FasL, TNFRII, RANTES, KGF, IGFI (AUC train: 0.95, AUC test: 0.89) and in Caucasians, TNF-α, MCP3, TGF-β3, TNFR1 and angiopoietin 2 (AUC train: 0.94 AUC test: 0.79). Conclusions Multivariate adaptive regression splines models multiple biomarkers associated with preterm birth and demonstrated racial disparity.