Predicting cardiac adverse events in patients receiving immune checkpoint inhibitors: A machine learning approach.

Abstract

e15075 Background: Many oncology treatments have been associated with cardiovascular (CV) adverse events. Cases of CV events, including myocarditis have been reported for PD-1 and PD-L1 therapies. We created a machine learning model to predict potential CV events in PD-(L)1 patients using the CancerLinQ database. Methods: A XGBoosted decision tree model was trained to predict a patient’s risk of serious CV adverse events. The model was trained on 80% of all advanced non-small cell lung cancer (NSCLC), melanoma, and renal cell carcinoma (RCC) patients from our database including those who received PD-(L)1 therapy. Index date was defined as date of first PD-(L)1 administration or date of advanced diagnosis if no PD-(L)1 drug was given. The model contained approximately 400 potential risk factors for cardiac disease including elements of past medical history, social history, vitals, common labs, cancer history (e.g. stage, cancer type), medication history, and PD-(L)1 specific factors including PD-(L)1 expression status and PD-(L)1 therapy administered. The model was tested on two separate validation sets (patients not used in training): one using advanced NSCLC, melanoma, and RCC patients and another with PD-(L)1 patients only. Each factor’s importance to the model’s predictions was calculated using SHAP summary plots, a qualitative technique for interpreting machine learning models. Results: A total of 27,172 advanced cancer patients were included in our study. 4,966 received PD-(L)1 therapy. The model was trained on 21,758 patients and 5,414 patients were set aside for testing. The model predicted serious cardiac events within 100 days of index with an AUC-ROC of 0.75 in all patients and 0.79 in PD-(L)1 patients. The top predictors of cardiac risk in PD-(L)1 patients included a history of heart disease, weight loss, the % lymphocyte count, and median LDH. The % lymphocyte count and weight loss were noticeably more predictive in PD-(L)1 patients than in non-PD-(L)1 patients. However, in general SHAP summary plots of all and PD-(L)1 patients were nearly identical, suggesting that both cohorts’ cardiac risk is determined in a similar way. PD-(L)1 and autoimmune disease associated factors did not appear in the top 40 most predictive risk factors. Conclusions: Using traditional cardiac risk factors, our model was able to predict potential cardiac events in PD-(L)1 patients. Our model found that high lymphocyte count may be protective while weight loss and a history of cardiac disease (e.g. heart failure) could indicate a poor prognosis.