Identifying Non-Linear Association Between Maternal Free Thyroxine and Risk of Preterm Delivery by a Machine Learning Model

OBJECTIVE: Preterm delivery (PTD) is the primary cause of mortality in infants. Mounting evidence indicates that thyroid dysfunction might be associated with an increased risk of PTD, but the dose-dependent association between the continuous spectrum maternal free thyroxine (FT4) and PTD is still not well-defined. This study aimed to further investigate this relationship using a machine learning-based model. METHODS: A hospital-based cohort study was conducted from January 2014 to December 2018 in Shanghai, China. Pregnant women who delivered singleton live births and had first-trimester thyroid function data available were included. The generalized additive models with penalized cubic regression spline were applied to explore the non-linear association between maternal FT4 and risk of PTD and also subtypes of PTD. The time-to-event method and multivariable Cox proportional hazard model were further applied to analyze the association of abnormally high and low maternal FT4 concentrations with the timing of PTD. RESULTS: A total of 65,565 singleton pregnancies with completed medical records and no known thyroid disease before pregnancy were included for final analyses. There was a U-shaped dose-dependent relationship between maternal FT4 in the first trimester and PTD (p <0.001). Compared with the normal range of maternal FT4, increased risk of PTD was identified in both low maternal FT4 (<11.7 pmol/L; adjusted hazard ratio [HR] 1.34, 95% CI [1.13–1.59]) and high maternal FT4 (>19.7 pmol/L; HR 1.41, 95% CI [1.13–1.76]). The association between isolated hypothyroxinemia and PTD was mainly associated with spontaneous PTD (HR 1.33, 95% CI [1.11–1.59]) while overt hyperthyroidism may be attributable to iatrogenic PTD (HR 1.51, 95% CI [1.18–1.92]) when compared with euthyroid women. Additionally, mediation analysis identified that an estimated 11.80% of the association between overt hyperthyroidism and iatrogenic PTD risk was mediated via the occurrence of hypertensive disorders in pregnancy (p <0.001). CONCLUSIONS: We revealed a U-shaped association between maternal FT4 and PTD for the first time, exceeding the clinical definition of maternal thyroid function test abnormalities. Our findings provide insights towards the need to establish optimal range of maternal FT4 concentrations for preventing adverse outcomes in pregnancy.