Prospective prediction of childhood body mass index trajectories using multi-task Gaussian processes

Arthur Leroy, Varsha Gupta, Mya Thway Tint, Delicia Shu Qin Ooi, Fabian Yap, Ngee Lek, Keith M. Godfrey, Yap Seng Chong, Yung Seng Lee, Johan G. Eriksson, Mauricio A. Álvarez, Navin Michael, Dennis Wang

TL;DR

This paper introduces MagmaClust, a novel probabilistic framework that uses multi-task Gaussian processes to model, cluster, and prospectively predict childhood Body Mass Index (BMI) trajectories. This approach offers a more robust and accurate method for identifying children at risk of future obesity, enabling earlier interventions.

What Problem Does This Paper Address?

Childhood obesity is a growing global concern, strongly linked to various health issues. While retrospective analysis of BMI trajectories has been studied, there's a significant gap in accurately prospectively predicting these trajectories and the associated risk of future obesity. Existing methods often struggle with missing data and lack the ability to provide reliable long-term forecasts. This makes it difficult for clinicians to proactively identify at-risk children and implement timely preventive measures. The paper aims to bridge this gap by developing a unified framework for modeling, clustering, and forecasting BMI trajectories with a focus on predicting future overweight/obesity risk.

Methods

The core of this work is the MagmaClust framework, which leverages multi-task Gaussian processes (GPs). GPs are probabilistic models that define a prior distribution over functions, allowing them to capture complex, non-linear relationships and quantify uncertainty. MagmaClust extends this by:

  • Multi-task Learning: Sharing information across individual BMI trajectories to improve modeling and clustering.
  • Clustering: Identifying distinct patterns (clusters) of BMI trajectories within the cohort.
  • Probabilistic Forecasting: Using historical BMI data to predict future BMI trajectories with associated uncertainty.
  • Non-parametric Approach: Avoiding rigid assumptions about the functional form of the trajectories.

MagmaClust was compared against two established methods: cubic B-splines (a non-parametric approach) and multilevel Jenss-Bayley models (a parametric, mixed-effects approach). The evaluation focused on robustness to missing data and accuracy in both retrospective reconstruction and prospective forecasting.

Datasets

The study utilized longitudinal BMI measurements from the Growing Up in Singapore Towards healthy Outcomes (GUSTO) cohort. This dataset includes BMI data for 1177 children from birth up to age 10 years. The data was split into a training set (N=600) and a test set (N=577) for model development and evaluation. No new benchmarks were introduced; rather, existing datasets were used to validate the proposed method against established techniques.

Key Insights and Ideas

  • The Power of Probabilistic Forecasting: MagmaClust's ability to provide a probability distribution for future BMI trajectories, rather than a single point estimate, is crucial. This allows for a nuanced understanding of risk and uncertainty, which is vital for clinical decision-making.
  • Robustness to Missing Data: The framework's superior performance with up to 90% missing data is a significant practical advantage, as real-world longitudinal datasets are often incomplete.
  • Early Childhood Trajectories Matter: The study highlights that distinct BMI patterns emerge very early in infancy (0-1 year), and these early differences have a significant impact on later BMI trajectories and obesity risk.
  • Unified Framework: MagmaClust offers a single, integrated tool for modeling, clustering, and predicting, simplifying the analysis pipeline compared to using separate tools for each task.

Key Results

  • Clustering: MagmaClust identified 5 distinct BMI trajectory patterns in children from birth to 10 years.
  • Missing Data Reconstruction: MagmaClust outperformed both cubic B-splines and Jenss-Bayley models in accuracy and robustness to missing data, with MSE being significantly lower (e.g., 24% lower than Jenss-Bayley and 79% lower than B-splines for 50% missing data).
  • Prospective Forecasting: MagmaClust demonstrated superior forecasting capabilities, with Jenss-Bayley's MSE being 1.62 to 3.26 times higher, and B-splines' MSE being 9 to 86 times higher, depending on the forecasting period.
  • Overweight/Obesity Prediction: Using BMI data from birth to age 2 years, MagmaClust achieved high specificity (0.94), negative predictive value (0.89), and accuracy (0.86) in predicting overweight/obesity status at age 10. These metrics improved with more data. For instance, sensitivity reached 90% when using data up to age 8 years.

Limitations and Future Work

  • Shrinkage Effect: The paper acknowledges that borrowing information from the population (shrinkage) might pull extreme individual trajectories towards the mean, potentially masking some unique growth patterns.
  • Data Specificity: While the GUSTO cohort is diverse, the model's generalizability to other ethnic groups or geographical locations might require retraining.
  • Future Work: The authors suggest that MagmaClust can be retrained with future growth data to capture post-pubertal trends and can be extended to incorporate more diverse populations. The framework's adaptability makes it suitable for ongoing assessments and enabling timely interventions. The ability to incorporate other environmental risk factors alongside BMI history is also a potential avenue for future research.

Summary updated

Abstract

MagmaClust provides a unified, probabilistic, non-parametric framework to model, cluster, and prospectively predict childhood BMI trajectories and overweight/obesity risk. The proposed method offers a convenient tool for clinicians to monitor BMI growth in children, allowing them to prospectively identify children with high predicted overweight/obesity risk and implement timely interventions.
BibTeX
@article{leroy2024prospective,
  author = {Arthur Leroy and Varsha Gupta and Mya Thway Tint and Delicia Shu Qin Ooi and Fabian Yap and Ngee Lek and Keith M. Godfrey and Yap Seng Chong and Yung Seng Lee and Johan G. Eriksson and Mauricio A. Álvarez and Navin Michael and Dennis Wang},
  title = {{Prospective prediction of childhood body mass index trajectories using multi-task Gaussian processes}},
  journal = {International Journal of Obesity},
  year = {2024},
  doi = {10.1038/s41366-024-01679-0},
}
Figure