A1012
Title: Clustering longitudinal data in clinical studies: A practical comparative analysis of statistical methods
Authors: Paola Rancoita - Vita-Salute San Raffaele University (Italy) [presenting]
Nicolo Pecorelli - Vita-Salute San Raffaele University (Italy)
Chiara Brombin - Vita-Salute San Raffaele University (Italy)
Abstract: Frequently, in clinical studies, patient status is monitored over time to decide patient management. These longitudinal data are challenging to analyze due to their intrinsic nature, irregular timing of measurements, missing values, and eventual nonlinearities. Moreover, often multivariate modeling (which jointly evaluates multiple outcomes) is needed to capture the multidimensional nature of the phenomenon. In this context, subject-specific heterogeneity frequently emerges, thus requiring the usage of robust statistical methods to identify latent patient subgroups, possibly related to their characteristics. This is particularly relevant for clinicians, as identifying clusters can allow the definition of personalized care strategies. From a statistical perspective, several clustering approaches for longitudinal data have been proposed, including algorithm-based (e.g., hierarchical clustering) and model-based approaches (e.g., finite mixture models, latent class mixed models). These methods differ in the number of cluster selection criteria, ability to include covariates or nested random effects, and to handle timing irregularities or complex trajectories. The flexibility and performance of these clustering strategies in uncovering latent patient longitudinal profiles will be compared. Their advantages and limitations are shown by applying them to real longitudinal data assessing quality of life and functional capacity in patients after pancreatic resection.
