A different perspective on studying stroke predictors: joint models for longitudinal and time-to-event data in a type 2 diabetes mellitus cohort

Abstract

Background Most predictive models rely on risk factors and clinical outcomes assessed simultaneously. This approach does not adequately reflect the progression of health conditions. By employing joint models of longitudinal and survival data, we can dynamically adjust prognosis predictions for individual patients. Our objective was to optimize the prediction of stroke or transient ischemic attack (TIA) via joint models that incorporate all available changes in the predictive variables. Methods A total of 3442 patients with type 2 diabetes mellitus (T2DM) and no history of stroke, TIA or myocardial infarction were followed for 12 years. Models were constructed independently for men and women. We used proportional hazards regression models to assess the effects of baseline characteristics (excluding longitudinal data) on the risk of stroke/TIA and linear mixed effects models to assess the effects of baseline characteristics on longitudinal data development over time. Both submodels were then combined into a joint model. To optimize the analysis, a univariate analysis was first performed for each longitudinal predictor to select the functional form that gave the best fit via the deviance information criterion. The variables were then entered into a multivariate model using pragmatic criteria, and if they improved the discriminatory ability of the model, the area under the curve (AUC) was used. Results During the follow-up period, 303 patients (8.8%) experienced their first stroke/TIA. Age was identified as an independent predictor among males. Among females, age was positively associated with atrial fibrillation (AF). The f inal model for males included AF, systolic blood pressure (SBP), and diastolic blood pressure (DBP), with albuminuria and the glomerular filtration rate (GFR) as adjustment variables. For females, the model included AF, blood pressure (BP), and renal function (albuminuria and GFR), with HbA1c and LDL cholesterol as adjustment variables. Both models demonstrated an AUC greater than 0.70.