Using a neural network to derive early childhood neurodevelopmental profiles in the New Hampshire Birth Cohort Study

Child-centred approaches represent a conceptual framework that emphasises the holistic characterisation of individual developmental patterns across cognitive, behavioural and social domains. As a complementary analytic tool, self-organising maps (SOMs), an artificial neural network, offer flexible, data-driven clustering capabilities that are well-suited to modeling complex, multidimensional and longitudinal developmental data. Despite their potential, few studies have applied such methods to profile early neurodevelopment, especially in rural populations.

We applied SOM to longitudinal neurobehavioural data (n=235) from healthy participant children from 3 to 5 years of age in the New Hampshire Birth Cohort Study, a rural pregnancy cohort. Group profiles reflect measures of behaviour and social responsiveness, cognition and motor performance and were examined in relation to known predictors of maternal-child characteristics using multinomial logistic regression as a proof of concept.

In our cohort, most children had neurotypical neurobehavioural scores, and 51% were boys. Mothers predominantly had some college education (74%), were married (93%) and were 31 years of age on average with above-average IQs relative to US norms. We identified six distinct neurobehavioural profiles (18-57 children each). The six profiles included: highest overall scores (profile 5), worst overall scores (profile 4), greatest behavioural/social improvement (profile 1), slight improvement (profile 3), average scores (profile 2) and highest adaptability (profile 6) relative to the full sample. Regression models showed expected associations with child sex, maternal IQ and parent-child relationships (eg, higher maternal IQ correlated with better cognitive outcomes).

Using a SOM, we identified distinct neurobehavioural profiles among rural children, reflecting variation across behaviour, social responsiveness, cognition and motor skills. These profiles varied by maternal and child characteristics and highlight the potential of neural network approaches to inform early risk or resilience identification in understudied populations.