Introduction

Congenital hypothyroidism (CH), an inborn thyroid hormone (TH) deficiency, is mostly caused by disturbances at the level of the thyroid (thyroidal CH, CH-T) or hypothalamus/pituitary (central CH, CH-C). Most CH newborn screening (NBS) programs are thyroid-stimulating hormone (TSH)-based, thereby only detecting CH-T. The Dutch NBS uses a multiple tier protocol. Decreased thyroxine (T4) concentrations may lead to measurement of TSH and thyroxine-binding globulin (TBG) aiming to detect both CH-T and CH-C, but at the cost of more false-positive referrals (positive predictive value (PPV) of 21%) compared to TSH-based screening. Recent studies describe a causative relation between THs and acylcarnitines (ACs) and amino acids (AAs) of which several are measured during NBS for other diseases. Therefore, we aimed to investigate whether ACs and AAs might contribute to discriminate newborns with CH from newborns with a normal CH screening and, with that, reduction of false-positive referrals.

Methods

Retrospective Dutch NBS data between 2007-2017 (gestational age and weight, T4, TSH, TBG, T4/TBG ratio, ACs and AAs) from 3436 newborns were used, including newborns with a normal CH screening (1842), false-positive referrals (1079) and newborns with CH-T (431) and CH-C (84). A Random Forest model including all these data was developed and its specificity, PPV and area under the curve (AUC) to predict CH were calculated.

Results

The Random Forest model to predict CH yielded a sensitivity of 100%, while obtaining a specificity of 83%, PVV of 51% and AUC of 0.99. Besides T4 and TSH, phenylalanine and tyrosine were the main factors contributing to the model’s performance. A second model that emphasized factors contributing to the prediction of CH-C showed that T4/TBG ratio contributed most, TSH did not contribute at all and succinylacetone was a new factor.

Discussion

The PPV of the Dutch NBS for CH may be improved from 21% to 51% by adding several ACs and AAs to a predictive machine learning based CH model. The next step will be testing the model in a prospective manner.