Multiclass classification of growth curves using random change points and heterogeneous random effects

V.w.l. Chin, Jarod Y. L. Lee, Louise Ryan, Robert Kohn, Scott A. SissonarXiv 2019 · DOI · arXiv

Abstract

Childhood growth faltering is a critical indicator of nutritional and environmental stress, often leading to long-term deficits in physical and cognitive development. Traditional "broken stick" models used to track these trajectories often rely on fixed knot locations and a predetermined number of subgroups, which can bias results. This paper introduces a flexible Bayesian multiclass classification model that incorporates child-specific random change points and a Dirichlet process prior to capture population heterogeneity. Applied to a longitudinal birth cohort from Vellore, India, the model identifies nine distinct growth subgroups, revealing significant correlations between early-stage faltering, maternal education, and later-life cognitive outcomes.

Introduction

Growth faltering—defined as a slower rate of growth compared to healthy reference populations—is a pervasive public health challenge in low-to-medium income countries. Beyond physical stature, faltering serves as a proxy for developmental stressors that impact intellectual and economic potential. Accurately identifying and characterizing these growth patterns is essential for designing targeted interventions. However, the statistical tools used to classify these trajectories often struggle with the inherent biological variability in the timing of growth milestones.

The Problem with Rigidity

Current statistical methods typically employ "broken stick" models—piecewise linear models where growth velocity changes at specific ages, known as knots. However, these traditional models suffer from two primary flaws:

  1. Rigid Timing: They assume every child experiences growth shifts at the exact same age (e.g., exactly at 6 months). If a child's milestones do not align with these fixed knots, the estimated growth velocity becomes biased.
  2. Fixed Subgroups: Researchers usually must pre-specify the number of growth patterns (subgroups) existing in a population. This often leads to models that either oversimplify the population or "hallucinate" extra subgroups to account for individual timing variations.

Proposed Methodology

To address these limitations, the authors propose a multiclass classification model that extends the broken stick framework through two key innovations:

Random Change Points

Instead of fixing knots at arbitrary ages, the model treats knot locations as child-specific random effects. This allows the "breaks" in the stick to shift based on when an individual child actually experiences a change in growth velocity. By accounting for individual differences in the onset of growth stages, the model provides a more accurate representation of individual trajectories.

Dirichlet Process (DP) Mixture Modeling

To solve the problem of determining the number of subgroups, the authors utilize a Bayesian non-parametric approach. The DP prior allows the number of mixture components (subgroups) to be data-driven. If the data is complex, the model adds groups; if it is simple, it collapses them. For inference, the researchers utilized Markov chain Monte Carlo (MCMC) algorithms, specifically a slice sampler, and applied the Posterior Expected Adjusted Rand (PEAR) method to finalize child classifications and resolve the common "label switching" problem in mixture models.

Evaluation and Key Results

The model was evaluated using both simulated data and real-world longitudinal data from the Vellore Birth Cohort (373 children in India), part of the Bill and Melinda Gates Foundation’s HBGDki project.

Simulation Insights

The simulation of 400 children with four distinct trajectories proved that the random change point model outperforms fixed-knot models. When knots are fixed incorrectly, models tend to create unnecessary subgroups to explain timing differences, complicating the public health interpretation. The random change point approach avoids this bias.

The Vellore Cohort Findings

When applied to the Vellore data, the model identified nine distinct subgroups of growth between birth and age one:

  • Subgroups 1 & 2: The most prevalent groups, showing relatively stable but slightly faltering trends.
  • Subgroup 3: Characterized by severe early-stage faltering in the first six months followed by a "catch-up" recovery phase.
  • Cognitive and Socioeconomic Links: The study found that children in subgroups 3 and 5 (severe early falterers) had lower median IQ scores (general, performance, and verbal) later in life compared to subgroup 4, which exhibited the mildest faltering. Additionally, maternal education was a strong predictor; mothers with no formal education were significantly more likely to have children in the more severe faltering subgroups.

Conclusion and Future Work

By allowing for individual-specific timing and data-driven subgroup identification, this model provides a more nuanced understanding of childhood growth. While the current application fixed the number of knots ($K=3$) for computational stability, future work aims to implement trans-dimensional approaches like Reversible-Jump MCMC to allow the number of knots to vary per child. Furthermore, scaling the model to the full HBGDki database will help determine if these nine subgroups remain consistent across different geographic and economic contexts.

Summary updated

Abstract

Faltering growth among children is a nutritional problem prevalent in low to medium income countries; it is generally defined as a slower rate of growth compared to a reference healthy population of the same age and gender. As faltering is closely associated with reduced physical, intellectual and economic productivity potential, it is important to identify faltered children and be able to characterise different growth patterns so that targeted treatments can be designed and administered. We introduce a multiclass classification model for growth trajectory that flexibly extends a current classification approach called the broken stick model, which is a piecewise linear model with breaks at fixed knot locations. Heterogeneity in growth patterns among children is captured using mixture distributed random effects, whereby the mixture components determine the classification of children into subgroups. The mixture distribution is modelled using a Dirichlet process prior, which avoids the need to choose the "true" number of mixture components, and allows this to be driven by the complexity of the data. Because children have individual differences in the onset of growth stages, we introduce child-specific random change points. Simulation results show that the random change point model outperforms the broken stick model because it has fewer restrictions on knot locations. We illustrate our model on a longitudinal birth cohort from the Healthy Birth, Growth and Development knowledge integration project funded by the Bill and Melinda Gates Foundation. Analysis reveals 9 subgroups of children within the population which exhibit varying faltering trends between birth and age one.
BibTeX
@misc{chin2019multiclass,
  author = {V.W.L. Chin and Jarod Y. L. Lee and Louise Ryan and Robert Kohn and Scott A. Sisson},
  title = {{Multiclass classification of growth curves using random change points and heterogeneous random effects}},
  year = {2019},
  eprint = {1909.07550},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/1909.07550},
}
Figure