Statistical learning to identify and characterise neurodevelopmental outcomes at 2 years in babies born preterm: model development and validation using population-level data from England and Wales

Abstract

Background: Children born preterm face elevated risks of neurodevelopmental impairments across domains. Prior studies have relied on expert-imposed typologies within single domains. This study applies statistical learning to a national database to identify transdomain clusters and their maternal and neonatal predictors.

Methods: Latent class analysis (LCA) was used to derive transdomain clusters from parent-reported visual, auditory, neuromotor, and communication impairments in preterm-born children at two years corrected age using the UK National Neonatal Research Database data (N = 27,261). Replication was conducted in an independent sample from Wales (N = 975). Clusters were clinically validated using cerebral palsy diagnosis, Bayley Scales of Infant and Toddler Development (3rd edition), and global neurodevelopmental delay. Random forest identified cluster-specific and shared predictors.

Findings: Four homogeneous clusters were derived (silhouette score = 0.71) and replicated in Wales with high balanced accuracy (93%): (1) typically developing (84.8%), (2) communication impairments (8.4%), (3) neuro-motor impairments (4.1%), and (4) multiple neuro-morbidity (2.7%). Clusters had high clinical validity and were distinguishable by shared and cluster-specific predictors. Neonatal brain injuries were most predictive of neuro-motor and multiple neuro-morbidity clusters. Birthweight, gestational age, socio-economic deprivation, and sex were stronger predictors of the communication cluster than preterm co-morbidities.

Interpretation: This study provides first evidence of the transdomain nature of neurodevelopmental impairments after preterm birth using LCA. The finding that socio-demographic and perinatal factors rather than co-morbidities increase the risk of communication impairment highlights the importance of environmental modification alongside clinical interventions. Applying data-driven approaches to routinely collected data may offer a cost-effective way to stratify at-risk children and inform targeted support strategies.

Funding: UKRI Medical Research Council.

Keywords: Birth cohorts; Machine learning; Neonatal; Neurocognitive; Neurodevelopmental impairments; Preterm.

Fig. 1

Overview of analytic steps. LCA was used to derive transdomain clusters of neurodevelopmental functioning using unvalidated parent-reported outcomes in the England cohort. Clusters were externally replicated using the Wales cohort. t-SNE algorithm was used to compare cluster structure in both cohorts. Clusters were clinically validated using three additional sources of data at the two-year follow-up, and RF was used to investigate shared and cluster-specific ante- and peri-natal features.

Fig. 2

Four clusters identified by latent class analysis in 27,261 children ina)England and 975 children inb)Wales at 2-years of age. The children were assigned to each cluster according to the maximum posterior probability. The lines display the prevalence of each impairment conditional upon cluster membership. TD = Typically developing, COMM = Communication impairments, NM = Neuro-motor impairments, MNM = Multiple neuro-morbidity.

Fig. 3

Two-dimensional representation of thea)England andb)Wales cohorts using t-SNE colour mapped by LCA clusters.

Fig. 4

Two-dimensional representation of thea)England andb)Wales cohorts using t-SNE colour mapped by impairments.

Fig. 5

Individual-level sequences stratified by a) LCA-derived clusters and b) healthcare professional global assessment of developmental delay (England). TD = Typically developing, COMM = Communication impairments, NM = Neuro-motor impairments, MNM = Multiple neuro-morbidity. For developmental delay (as diagnosed by a HCP), typically developing: <3 months delay; mild delay: 3–6 months delay; moderate delay: 6–12 months delay; severe delay: >12 months delay. Severity of impairments defined in Table 1. 35% of children assigned to MNM had 5–8 impairments compared with 19% who were assessed as severely delayed.

Fig. 6

Alluvial plot to show change in individual assignments between HCP assessment of developmental delay and LCA cluster assignment (England cohort). This plot shows that there was low agreement in assignments between both classifications among children with impairments. For example, children from all four developmental delay categories were assigned to the TD cluster. Only 41% with severe developmental delay were assigned to MNM and 10% to TD. N = 26,032 children with both LCA assignment and HCP assessment of developmental delay. Adjusted Rand Index: 0.457.

Fig. 7

Global interpretation of the contribution of the top fifteen features to the model predictions of each cluster (England cohort). The bar plots show the mean absolute SHAP value of each feature across all observations per cluster in the England cohort. The bars show the impact each feature has on the model predicting each cluster from the RF model. The order of the bars shows the contribution (feature importance) each predictor makes in determining the output of the models (top to bottom = highest to lowest contribution).