Developing A Model to Predict Major Bleeding Among Hospitalized Patients Undergoing Therapeutic Plasma Exchange

Abstract

Although therapeutic plasma exchange (TPE) can be associated with bleeding, there are currently no known strategies to reliably predict bleeding risk. This study developed a TPE bleeding risk prediction model for hospitalized patients. To develop the prediction model, we undertook a secondary analysis of public use files from the Recipient Epidemiology and Donor Evaluation Study-III. First, we used a literature review to identify potential predictors. Second, we used Multiple Imputation by Chained Equations to impute variables with < 30% missing data. Third, we performed a 10-fold Cross-Validated Least Absolute Shrinkage and Selection Operator to optimize variable selection. Finally, we fitted a logistic regression model. The model identified 10 unique predictors and seven interactions. Among those with the highest odds ratios (OR) were the following: > 10 TPE procedures and antiplatelet agents (OR 3.26); nephrogenic systemic sclerosis (OR 3.15); and intensive care unit stay (OR 3.08). Among those with the lowest OR were the following: albumin-only TPE (OR 0.50); male sex (OR 0.82); and heart failure (OR 0.85). The model indicated an acceptable performance with a C-statistic of 0.71 (95% CI 0.699-0.717). A model to predict bleeding risk among hospitalized patients undergoing TPE identified key predictors and interactions. Although the model achieved acceptable performance, future studies are needed to validate and operationalize it.