(C-252) Development and Validation of a Machine Learning Model to Predict Rituximab-Induced Infusion Reactions in B-Cell Malignancy

Background: High proportion of patients who use rituximab (RTX) to treat B-cell malignancy experience infusion-related reactions (IRRs) during their first cycle of RTX administration.

Objectives: We developed and validated a machine learning (ML) model to predict the occurrence of RTX related IRRs.

Methods: This study was a retrospective case-control study using multi-institutional electronic medical records (EMRs) and clinical data warehouses (CDWs). It targeted B-cell malignancy patients over the age of 18 who first used RTX from January 1, 2009, to December 31, 2022, at tertiary academic hospitals. For model development, patient data from Seoul National University Hospital was used. We identified cases as patients who experienced RTX-induced IRRs during the first cycle, and the control group was randomly selected from patients who did not experience IRR while receiving rituximab during the first cycle as the same number of patients as the case group. Variables were selected based on several literatures, and retrospectively collected. The entire data was divided into a train set and an internal validation set in a ratio of 8:2. To find the optimal combination of features for developing the ML prediction model, recursive feature elimination with the cross-validation (RFECV) method was used. The prediction models were developed using the four ML algorithms: Random Forest, eXtreme Gradient Boosting (XGBoost), Support Vector Machine, and Polynomial linear regression. The model with the highest area under the receiver operating characteristic curve (AUROC) among ML algorithms was chosen. To interpret the significance and direction of the selected variables in the optimal model, Shapley Additive exPlanations (SHAP) values were used. For external validation to evaluate the reproducibility and generalizability of the final selected model, patient data from Seoul National University Bundang Hospital was collected under the same criteria as in model development and used as the external validation set.

Results: For model development, we collected the data from 592 cases and 592 controls. With seven variables selected in RFECV, the XGBoost model showed the highest performance (AUROC = 0.71; PPV = 0.65; NPV = 0.64). Age, lymphocyte count, and corticosteroid premedication were the three most important variables for predicting IRR. For external validation, we collected the data from 121 cases and 121 controls. The XGBoost model also showed consistent performance in external validation results (AUROC = 0.75; PPV = 0.78; NPV = 0.62).

Conclusions: We developed and validated an optimal machine learning model to reproducibly predict the occurrence of IRR with the first dose of rituximab in Korean patients with B-cell malignancies.