A personalized prediction of longitudinal growth using People-Like-Me methods

Conventional prediction models typically emphasize overall trends, often missing the variability in individual responses, which limits their effectiveness for personalized predictions. The People-Like-Me (PLM) methodology was developed to address this issue by employing curve-matching techniques to generate individualized predictions. PLM is a data-driven algorithm that selects similar matching trajectories to estimate the trajectory for an out-of-sample target. In this study, we extend the PLM methods by introducing the Mahalanobis distance as a new metric for selecting matches, allowing for the consideration of correlations between time points in longitudinal data. We assess the performance of this enhanced PLM across various scenarios using clinical growth data from children with cystic fibrosis and simulated datasets. Our analysis compares (i) different match selection strategies, and (ii) PLM predictions with those from linear mixed models (LMM). The results consistently show that Mahalanobis-based PLM outperforms both the standard PLM and LMM. This establishes Mahalanobis-based PLM as a more accurate and flexible method for personalized prediction of longitudinal trajectories.