Identifying longitudinal-growth patterns from infancy to childhood: a study comparing multiple clustering techniques

Background: Most studies on children evaluate longitudinal growth as an important health indicator. Different methods have been used to detect growth patterns across childhood, but with no comparison between them to evaluate result consistency. We explored the variation in growth patterns as detected by different clustering and latent class modelling techniques. Moreover, we investigated how the characteristics/features (e.g. slope, tempo, velocity) of longitudinal growth influence pattern detection. Methods: We studied 1134 children from The Applied Research Group for Kids cohort with longitudinal-growth measurements [height, weight, body mass index (BMI)] available from birth until 12years of age. Growth patterns were identified by latent class mixed models (LCMM) and time-series clustering (TSC) using various algorithms and distance measures. Time-invariant features were extracted from all growth measures. A random forest classifier was used to predict the identified growth patterns for each growth measure using the extracted features. Results: Overall, 72 TSC configurations were tested. For BMI, we identified three growth patterns by both TSC and LCMM. The clustering agreement was 58% between LCMM and TS clusters, whereas it varied between 30.8% and 93.3% within the TSC configurations. The extracted features (n=67) predicted the identified patterns for each growth measure with accuracy of 82%-89%. Specific feature categories were identified as the most important predictors for patterns of all tested growth measures. Conclusion: Growth-pattern detection is affected by the method employed. This can impact on comparisons across different populations or associations between growth patterns and health outcomes. Growth features can be reliably used as predictors of growth patterns.

Automated summary

This study examined the variation in growth patterns detected by different clustering and latent class modelling techniques, as well as how characteristics of longitudinal growth influence pattern detection. Results showed that the method employed can impact growth-pattern detection and that growth features can be reliably used as predictors of growth patterns.