Combining human milk from several donors reduced nutrient variation by 39%

Summary

• Using historical data from 2,236 milk pools contributed by 480 donors, the multi-donor strategy reduced average deviation from target protein and energy levels by 39% compared with single-donor pooling.

  
  

• The model combined milk from three to five donors into two-litre pools while accounting for nutrient targets, available volume and expiry dates.

  
  

• The study is a proof of concept: the proposed pools were tested on historical data, so the next step is to prepare them in a milk-bank setting and measure their actual nutrient content.

  
  

Why can the nutritional content of donor human milk vary?

Human milk naturally changes between people and across the months after birth. Protein and energy content can therefore differ from one donated batch to another.

For premature infants, consistent nutrition is important because growth and clinical needs are closely monitored. Many European milk banks pool milk from one donor at a time, which preserves straightforward traceability but means nutrient content can vary between successive pools.

Why is this a prescriptive analytics problem?

A milk bank has many individual bottles in storage and must decide which ones to combine into each pool. That decision has several goals at once: reach target protein and energy levels, create the required volume, use milk before it expires and follow safety and traceability rules.

The researchers combined prediction with optimisation. Machine learning first estimated the likely nutrient content of available milk. An optimisation model then selected which bottles to combine so that each pool came as close as possible to the nutrition targets.

How did the collaboration build and test the approach?

The project brought together Amsterdam UMC's Dutch National Human Milk Bank, the University of Amsterdam and the University of Toronto.

The team trained machine-learning models using donor information already collected by the milk bank, including age, body mass index, diet, time since giving birth and expressed milk volume. A random-forest model - a method that combines many decision trees to make a prediction - performed best among the approaches tested.

Those predictions were used in an optimisation model that formed two-litre pools from three to five donors. The analysis used historical data from 2,236 milk pools contributed by 480 donors between 2017 and 2024.

What did the findings show?

Compared with the existing single-donor approach, the multi-donor pooling strategy reduced the average deviation from target protein and energy levels by 39% in the historical-data analysis.

Time since giving birth was the strongest predictor of milk composition in the model, followed by donor body mass index and the amount expressed per pumping session. The team also observed a difference in protein content between vegetarian and omnivorous donors, but this finding requires further study.

Multi-donor pools would still be pasteurised and fully documented, so safety and traceability would remain in place. The main operational trade-off is shelf life: the whole pool must be used by the earliest expiry date of any bottle it contains. The study also did not assess infant growth or other clinical outcomes. The 39% improvement therefore refers to more consistent nutrient content, rather than demonstrated health benefit.

How can the findings inform other applications?

The next step is prospective testing in a milk bank: preparing pools selected by the model, measuring their actual protein and energy content and assessing how the approach fits routine operations.

More broadly, the study showed how prediction and optimisation can be combined when a health service needs to blend variable biological materials to meet a target while respecting safety, inventory and expiry constraints.