Development and external validation of machine learning algorithms for postnatal gestational age estimation using...

Area of Science:

• Biomedical data science
• Newborn health research
• Metabolomics applications

Background:

• Accurate gestational age (GA) estimation is crucial for preterm birth surveillance, especially in low-resource settings.
• Traditional methods for GA assessment can be challenging in low-income countries.
• Machine learning offers a potential solution for precise GA estimation shortly after birth.

Purpose of the Study:

• To develop and validate machine learning models for accurate gestational age estimation.
• To utilize clinical and metabolomic data for GA prediction.
• To assess model performance in diverse, external populations.

Main Methods:

• Developed three GA estimation models using ELASTIC NET multivariable linear regression.
• Utilized metabolomic markers from heel-prick blood samples and clinical data.
• Validated models internally in Ontario, Canada, and externally in Zambia and Bangladesh.

Main Results:

• The best-performing model estimated GA within approximately 6 days of ultrasound in Zambian and Bangladeshi heel-prick blood samples.
• Model accuracy was slightly lower with cord blood samples, estimating GA within about 7 days.
• Mean absolute error (MAE) was below 0.81 weeks for heel-prick data in external cohorts.

Conclusions:

• Machine learning algorithms developed in Canada accurately estimated gestational age in external cohorts from Zambia and Bangladesh.
• The models demonstrated superior performance when using heel-prick blood data compared to cord blood data.
• This approach shows promise for improving preterm birth surveillance in resource-limited settings.