• Rachel Wong

Human milk banking: Can machine learning predict the macronutrient content of incoming donations?

Updated: May 20


Picture of human milk in collection bottles donated to a milk bank
Human milk donated to a milk bank. Source: United States Breastfeeding Committee.

Human donor milk varies in macronutrient content due to many factors such as lactation stage, infant gestational age, time of day when pumping, and volume expressed. Therefore, human milk banks usually combine several donations into a batch to increase the immunological variety and decrease the variation in the macronutrient composition of the milk fed to babies in the neonatal intensive care unit (NICU). However, there is no agreed-upon method to select which donations to pool for optimal content since donations' macronutrient content is unknown upon arrival.


Babies in the NICU, particularly those born with a very low birth weight (VLBW), have high nutrient needs to support their growth. Variability in nutrient composition of donated milk complicates the production of a uniform pooled product and, subsequently, the provision of adequate nutrition to promote optimal growth and development of VLBW infants. Given the high nutrient needs of these vulnerable babies, exploring ways to optimize the recipe for combining donations is essential.


What did we do?

In this study, our former undergraduate research student Rachel Wong aimed to understand whether we could use features like mother's pumping practices and mother and infant characteristics to predict the macronutrient content of donations at the milk bank. We focused on predicting fat and protein, given their well-established importance in the growth of very-low-birth-weight infants.


Over 12 months, we worked with the Rogers Hixon Ontario Human Milk Bank to collect donation features identified to influence fat or protein content in some way. These features include: the time of day the mother pumped the milk (morning, afternoon, night), average volume donated per day, donor age, infant gestational age, and the number of days between the first pump and last pump dates.


We applied several common machine learning models to this set of features to answer the following questions:

  1. Can different machine learning models learn to make accurate predictions?

  2. What features were most predictive of fat and protein content?


 

You can read the study published in the Journal of Nutrition here: https://academic.oup.com/jn/article-abstract/151/7/2075/6224883

 

What did we find?
  • Using donor information typically collected by milk banks, common machine learning models can accurately predict each donations' fat and protein content within a reasonable margin of error.

  • The most predictive features included whether the infant was born preterm vs. term and the lactation stage of the donor.


What are the next steps?

The ability to successfully predict the fat and protein content of donations that arrive at the milk bank allows us to use this information to select the donations to combine to maximize the nutrient composition of milk fed to infants born VLBW.


We are currently looking to propose a two-step predict-optimize model to examine this process. First, we use a machine learning model to predict the macronutrient values of all available donations. Second, based on the first step's macronutrient predictions, we apply an optimization model to decide which donor deposits to pool together in a recipe.



 

To learn more about human milk banking, or to become a donor, visit: https://www.milkbankontario.ca/