Development and validation of a dynamic prediction model for infective endocarditis patients based on platelet trajectories latent classes: a longitudinal retrospective study from multicenter databases

Background Platelets (PLTs) are a driving factor in infective endocarditis (IE) and, as the smallest cellular component in the blood, are sensitive to the effects of infection status, immune status, and the degree of frailty. IE, a disease resulting from the cumulative convergence of multisystem mechanisms involving infection, hemodynamics, immunity, and coagulation that locally target the cardiac endothelium, demonstrates marked heterogeneity in both pathogenic manifestations and clinical outcomes. Therefore, the aim of this study was to explore PLT trajectories during treatment and the clinical characteristics of IE patients of different trajectories. Methods We conducted a retrospective analysis of longitudinal data from multiple databases (eICU and MIMIC). Latent class growth mixture modeling (LCGMM) was implemented to identify PLT trajectories and perform cluster analysis. Model selection criteria [log-likelihood, Akaike information criterion (AIC), Bayesian information criterion (BIC), and entropy] and average posterior probabilities were calculated to determine the optimal number of trajectory classes. Cox proportional hazards and logistic regression analyses were conducted to evaluate associations between trajectory subgroups and clinical outcomes. Bayesian joint models were subsequently developed to construct dynamic prediction models, with model parameters estimated using Markov chain Monte Carlo (MCMC) algorithms. The predictive performance of the dynamic models was assessed through the area under the receiver operating characteristic (AUROC) curve at multi-timepoint. Results Through clustering analysis of IE cohorts on PLTs post-admission, we identified four latent classes, each exhibiting unique clinical profiles (entropy: 0.815). We established a dynamic predictive model which integrates infection status (blood culture and white blood cell) and PLTs updated with each test during ICU stay, and achieved robust predictive performance (AUC: 0.71, Youden: 0.96, F1-score: 0.95). Conclusion Integrating longitudinal PLTs trajectories with baseline characteristics enables effective risk stratification and adverse outcome prediction in patients with IE. Graphical abstract Illustrative schematic depicting the workflow and findings of this study (created with BioRender.com). First, we identify latent classes among IE patients based on a latent class growth mixed model. Subsequently, we incorporate the identified latent classes as predictors into the survival sub-model, constructing a Bayesian joint model by combining longitudinal nonlinear models and non-Gaussian joint models. IE: infective endocarditis, LCGMM: latent class growth mixed model, AUC: area under curve.