Figure 3

Aggregation of nutritional clusters based on admission variables identifies 9 admission groups to further assess association with geographical location and outcome variables. Hierarchical clustering and PCA performed on the 24 most populated nutritional clusters (covering 95% of the whole cohort) using admission variables (shown in (a)) to characterize infants in each cluster. Nutritional clusters indexed according to size rank. The 9 admission groups, defined by visual inspection, are encoded with letters A to I on the hierarchical clustering dendrogram (a) and on the PCA projection plot (b). (a) Admission z-scores values (used for clustering), along with outcome variables (z-scores), nutritional and geographical location variables (mean proportions, Table 2). Colour coding for admission and outcome variables uses green for “favourable” and purple for “unfavourable” values (e.g. a high z-score value is unfavourable for NEC while favourable for MM at discharge). Feature columns in admission, outcome and geographical location variables are also reordered based upon hierarchical clustering. (b) Projection of the nutritional clusters (plotted as points with cluster index as in (a)) along the two main principal components returned by PCA (PC1 and PC2, explaining 94% of the variance). Admission groups A to I from (a) are reproduced on the PCA plot with ellipses whose axes are based on the covariances of the group’s coordinates along PC1 and PC2. (c) Contributions of individual admission variables to the first two principal components (PC1 and PC2). PC1 is driven by gestational age, resuscitation and Apgar scores, while PC2 is driven by gestational age and birth weight z-score.