Can machine learning predict school-age language from maternal diet and family factors?

A feasibility test, not a population predictor

The paper asks whether multivariate machine learning can link what mothers ate during lactation, together with family and child characteristics, to neurocognitive scores at early school age. Its answer is a qualified yes: the approach is feasible and produces candidate predictors, but the authors state that no specific conclusions can be drawn about the role of these factors in the general population of term-born infants.

The sample is 20 mother-child dyads drawn from an existing cohort of 160 typically developing late-preterm and full-term neonates recruited in Boston between 2014 and 2017. Seventeen of the 20 were breastfeeding and three were formula-fed. The authors attribute the small analyzed sample partly to COVID-19 disruptions that caused loss of follow-up, and they report that the 20 included participants were broadly similar to the larger cohort on gestational age, sex, birth weight, birth length, head circumference, maternal age, maternal education, family income, maternal income, and feeding mode.

Diet from a questionnaire, language from parents

Maternal diet was captured with a self-administered 139-item Block 2005 Food Frequency Questionnaire covering the past three months, including meals, snacks, and supplements. Standardized software converted responses into 73 nutrient measurements grouped into six categories: total energy, carbohydrates, fats, minerals, vitamins and antioxidants, and fibers. These nutrient values, not foods, became the dietary inputs to the model.

Neurocognition was measured with the Child Development Inventory, a parent questionnaire for children roughly 15 months to 6 years old. Its 270 developmental statements form eight scales, including expressive language, language comprehension, letters, and numbers. Children were on average 3.86 ± 0.65 years old when the inventory was completed. This matters for reading the results: the outcome is a parent-reported developmental score, and the paper's "prediction error" refers to how closely the model reproduced those scores.

Single-variable analysis found almost nothing

Before any machine learning, the authors tested each predictor against each outcome one at a time, using Spearman correlation for continuous variables and ANOVA for binary ones, with age as a covariate and false discovery rate correction for multiple comparisons. After that correction, no non-nutritional variable was associated with any inventory score. The only significant univariate association was between maternal sugar intake and the self-help domain (corrected p = 0.0163).

That near-empty univariate result is the setup for the paper's main methodological claim. The authors argue that cohort-level associations between one variable and one outcome do not necessarily explain individual differences, so they turned to a model that selects combinations of features for each child.

Feature selection keeps panels smaller than the sample

The model used Sequential Forward Floating Selection, which starts with no features, adds the one that most reduces prediction error, then tests whether removing an already-selected feature improves things further, alternating these steps until no change helps. Prediction error was measured as mean absolute percentage error, chosen because it scales errors to each score's own range and so allows comparison across domains.

From 86 candidate features, the pipeline selected 6 to 16 per outcome—fewer features than the 20 participants, a deliberate guard against overfitting in a small, high-dimensional dataset. The authors also report using repeated runs, comparison with baseline models, bootstrap confidence intervals, and permutation testing, while acknowledging these cannot fully eliminate overfitting risk in a cohort this size.

Expressive language was easiest to predict

For expressive language, the selected panel combined non-nutritional variables—family income and birth length—with nutritional ones including carbohydrate, energy, cholesterol, phosphocholine, total and reduced glutathione, glycemic load, maltose, and sucrose. The abstract lists phosphatidylcholine, glutathione, glycemic load, and other nutrients alongside child demographics and family income as candidate predictors; the full-text excerpt names phosphocholine and parental education levels. These are the features the algorithm retained, not established causes.

Prediction error was lowest for expressive language at 3% and highest for letter scores at 72%. The authors describe the association strength as greatest for expressive language and least for letter scores. Because the metric is percentage error against each child's actual score, the gap means the model tracked expressive language scores closely and letter scores poorly—it does not mean language is more biologically determined.

Combining diet and family factors beat either alone

The paper's principal finding is that combining maternal nutrient intake with sociodemographic factors predicted early school-age neurocognitive outcomes better than either factor alone. Features frequently selected across different outcomes included highest maternal degree and birth length among non-nutritional variables, and DPA and DHA among nutritional ones. The self-help panel, by contrast, drew on gestational age, birth length, maternal education, EPA, DPA, DHA, copper, sugar, lactose, and sucrose.

The authors frame this as a correction to prior work that studied single nutrients in isolation or set sociodemographic variables aside. Their interpretation is that nutrients interact with each other and with environmental factors, and that diet is potentially more modifiable than genetics, birth demographics, income, or education. That modifiability argument is the authors' rationale for why the combination is worth studying, not a demonstrated intervention effect.

Why 3% error is not a clinical claim

A 3% mean absolute percentage error on expressive language in 20 children is a statement about how well a fitted model reproduced scores in a very small sample, evaluated with leave-one-out runs and resampling checks. It is not evidence that a diet-and-demographics model would predict a new child's language with 3% error, and the authors explicitly limit conclusions to feasibility and hypothesis generation.

The measurement choices also bound the claim. Diet came from a food frequency questionnaire about the prior three months, converted to nutrients by commercial software; the outcome came from a parent-completed inventory. Neither is an objective biomarker or a clinician-administered test, and the paper does not report a held-out external validation cohort.

The useful signal is the domain ranking

The most decision-relevant result in this paper is not the 3% figure by itself but the ordering across domains: expressive language was predicted with the least error, letter scores with the most, and the selected feature panels differed by outcome. For readers weighing whether to pursue nutrition-brain modeling, that pattern says the approach can generate outcome-specific candidate predictor sets worth testing at larger scale, and that literacy-related letter knowledge is the weakest target in this sample.

The authors position the work as an impetus for larger studies of nutrition-brain interactions that could support hypothesis generation for individualized nutritional recommendations during pregnancy and lactation. The pilot's contribution is therefore a prioritized, testable feature list—maternal diet during lactation plus family income and parental education—rather than a predictive tool ready for use.

Sources used in this answer