Identification of intradialytic hemodynamic phenotypes using latent profile analysis

Intradialytic blood pressure variations are frequent and associated with poor clinical outcomes. However, the use of arbitrary, inconsistent diagnostic definitions limits cross-study comparability and hinders the development of evidence-based knowledge guiding management. This study aimed to bypass rigid, threshold-based diagnostic criteria using latent profile analysis, an unsupervised clustering method, to identify data-driven hemodynamic phenotypes in incident hemodialysis patients and explore their prognostic utility.

We applied latent profile analysis to intradialytic hemodynamic time-series data from a retrospective cohort of incident maintenance hemodialysis patients from 2017 to 2022. Sixteen patient-level indicators capturing baseline hemodynamics were manually engineered. We established internal validity via cross-validation, identified baseline clinical predictors of profile membership, and assessed prognostic utility against standard diagnostic definitions using Cox proportional hazards models.

A four-profile model was selected based on fit diagnostics and clinical interpretation and demonstrated cross-validation robustness. Profile 1 exhibited early nadirs with a late-session recovery. Profile 2 displayed a highly oscillatory pattern. Profile 3 demonstrated remarkable hemodynamic stability. Profile 4 featured frequent, unrecovered early nadirs. While results from the survival analyses were non-significant, profiles yielded divergent point estimates and demonstrated comparable discrimination and statistical fit to standard definition-based models.

Unsupervised clustering effectively uncovers distinct hemodynamic phenotypes, bypassing the limitations of rigid diagnostic thresholds. Data-driven profiling provides a rigorous taxonomy to improve personalized clinical decision-making and reduce heterogeneity in future interventional studies.