P1533Dealing with missing patient characteristics in clinical practice when using cardiovascular prediction models

Abstract

Background: Cardiovascular prediction models can be used to increase the efficiency of preventive therapy. In clinical practice there are often one or more missing values for risk factors and this may be one of the reasons why prediction models are underused. Simpler models with fewer variables are generally less accurate and of limited value and, a practical and flexible approach for dealing with missing values is needed. Purpose(s): To compare the validity of pragmatic approaches for dealing with missing patient characteristics in cardiovascular risk prediction. Method(s): As an example, we used real-world data from a registry of patients >18 years old with type 2 diabetes mellitus (T2DM) included in the Swedish National Diabetes Registry (NDR; n=419,533). First, an updated version of the 5-year NDR risk equation to predict CVD was developed in a random 25% of patients. The following risk factors (percentage missing at baseline) were included: age (0%), sex (0%), age at onset of T2DM (12%), smoking status (22%), BMI (26%), systolic blood pressure (15%), HbA1c (11%), non-HDL-cholesterol (40%), albuminuria (42%), eGFR (21%), retinopathy (67%), and a history of cardiovascular disease (0%) and atrial fibrillation (0%). In addition to the risk equation, the 25% sample was also used to generate the framework for 3 methods for handling missing data. 1) Reduced model method. Starting with a model including all predictors, all possible models with a combination of fewer predictors were developed, resulting in 512 different prediction models. Although not very practical due to the numbers of models, theoretically, the reduced model method is the gold standard. The models with fewer predictors are the best fitting models in the situation of unavailable data. 2) Conditional single imputation to impute variables based on the available characteristics. All missing values are estimated with one linear or one logistic regression model in R for continuous and categorical predictors respectively. In the case of multiple missing variables, imputation consists of 30 iterations. 3) Median imputation method, imputing median values for continuous predictors and mean values for categorical predictors from the development dataset. Next, the risk score was validated in the remaining 75% of the patients from the Swedish NDR in the presence of missing characteristics, using different approaches to deal with missing data. Result(s): The reduced model method, single imputation method, and median imputation method all had good agreement between predicted and observed risks. C-statistics were comparable between 0.81 and 0.84 for any number of missing patient characteristics (figure 1). Conclusion(s): Missing patient characteristics should not stop physicians from pre-dicting CVD risk in clinical practice. If important characteristics such as age, sex and history of CVD are available, pragmatic imputation of median values results in reliable predictions.