Development and external validation of machine learning algorithms for postnatal gestational age estimation using clinical data and metabolomic markers

Abstract

Accurate estimates of gestational age at birth are important for preterm birth surveillance but can be challenging to reliably acquire in low and middle income countries. Our objective was to develop machine learning models to accurately estimate gestational age shortly after birth using clinical and metabolic data. We derived and internally validated three models using ELASTIC NET multivariable linear regression in heel prick blood samples and clinical data from a retrospective cohort of newborns from Ontario, Canada. We conducted external model validation in heel prick and cord blood sample data collected from prospective birth cohorts in Lusaka, Zambia (N=311) and Matlab, Bangladesh (N=1176). The best-performing model accurately estimated gestational age within about 6 days of early pregnancy ultrasound estimates in both cohorts when applied to heel prick data (MAE (95% CI) = 0.79 weeks (0.69, 0.90) for Zambia; 0.81 weeks (0.75, 0.86) for Bangladesh), and within about 7 days when applied to cord blood data (1.02 weeks (0.90, 1.15) for Zambia; 0.95 weeks (0.90, 0.99) for Bangladesh). Algorithms developed in Canada provided accurate estimates of gestational age when applied to external cohorts from Zambia and Bangladesh. Model performance was superior in heel prick data as compared to cord blood data.

Unlocking the global health potential of dried blood spot cards

“…These samples contain information on the metabolic "fingerprint" of infants which can be used to identify whether they are at risk for specific inborn errors of metabolism and other genetic conditions. We previously demonstrated that data from DBS-derived metabolites can also be utilised to accurately estimate gestational age [2][3][4][5][6]. Gestational age dating is challenging in LMICs due to the lack of access to prenatal ultrasound and the unreliability of dating based on the last menstrual period [7,8].…”

“…Sample receipt and analysis were integrated into the existing NSO workflow. Further details on analytes measured can be found here [4,5,10]. Priority analysis times were put in place for all biochemical testing and hemoglobinopathy screening, with the goal of completing analysis of priority conditions within 14 days of birth.…”

A novel way of determining gestational age upon the birth of a child

“…Another challenge is the adaptation of algorithms developed using blood spot data to estimate population rates of SGA. Existing algorithms can be enhanced using machine learning approaches to improve model performance, and could be further reiterated for SGA-specific estimations [ 11 ].…”

Metabolic gestational age assessment in low resource settings: a validation protocol

“…In the Bangladeshi cohort, the model correctly estimated ultrasound-validated gestational age to within one week in the majority of infants, and within two weeks in 90-95% of infants using heel prick and cord blood metabolomic markers 6 . We have since developed and internally validated a second generation of models using a machine learning algorithm and have conducted external validation of model performance in newborn data from two cohorts from Zambia and Bangladesh 7 . The successful use of cord blood in estimating gestational age is of particular interest as we encountered barriers to heel prick sample collection in these populations.…”

“…While our previous work, and that of others 5,7 , suggests potential value in using dried blood-spot-derived analytes for gestational age assessment, some key questions need to be answered before attempting to utilize this approach widely as a surveillance tool for preterm birth:…”