Artificial intelligence surrogate models to predict long-term cardiovascular effects of immune checkpoint inhibitor therapies using electrocardiograms.

12020 Background: Immune checkpoint inhibitors (ICI) revolutionized the treatment landscape for many cancers, with close to 50% of cancer patients now ICI eligible. Acutely, ICI are associated with myocarditis. Long term cardiovascular effects of ICI are less clear. We built artificial intelligence (AI) models as surrogates for predicting long-term cardiovascular disease (CVD) after ICI using electrocardiograms (ECG). Methods: Using data from over 80,000 cancer patients treated from 1986 to 2021 at the University of California, San Francisco, we develop two ECG AI models as surrogates for risk over time of CVD. First, we built a model of observed CVD using true outcomes in a time to event framework. This model is a surrogate for prevalence under the current standard of care. Second, we build a model for causal outcomes using before-and-after data and estimate risk from ECGs for causal analyses. We hold out patients (N=15,277), including all those treated with anthracyclines (N= 3,681), trastuzumab (N=751), or ICI (N= 3,572), for evaluation. Causal effects are estimated with paired pre and post treatment ECGs within three years. Results: Models have average AUC across CVDs and years of 0.78 and 0.77. The observed model estimates ICI treated patients compared to those with any other cancer treatment have higher average 10-year risk of ischemic heart disease (IHD) by 10% (95% CI: 2-18%), venous thromboembolism (VTE) by 10% (1-20%), and critical ventricular arrhythmias (CVA) by 10% (3-19%), largely due to high baseline risk. Anthracycline and trastuzumab treated patients did not have higher 10-year heart failure (HF) risk relative to other treatments in aggregate. The causal framework estimates anthracyclines increased average 10-year risk of HF by 45% and trastuzumab by 46%. Anthracyclines and trastuzumab significantly increased risk of atrial fibrillation, IHD, ischemic stroke, CVA, VTE, and conduction disorders as well. ICI increased risk of each of these significantly by smaller amounts. Notably, after ICI, average 10-year risk of HF increased by 32% and IHD by 17%. Conclusions: Our study demonstrates potential long-term CVD effects of ICI as estimated from ECGs. This framework can be used to evaluate effects of new therapies in the future. 10-year pre/post treatment relative risks. CVD Anthracycline Trastuzumab ICI Atrial fibrillation 1.34 [1.29, 1.39] 1.30 [1.24, 1.36] 1.25 [1.21, 1.28] Ischemic heart disease 1.25 [1.21, 1.28] 1.28 [1.23, 1.32] 1.17 [1.14, 1.19] Heart failure 1.45 [1.39, 1.52] 1.46 [1.38, 1.55] 1.32 [1.27, 1.36] Ischemic stroke 1.41 [1.32, 1.50] 1.31 [1.22, 1.41] 1.28 [1.21, 1.35] Critical ventricular arrhythmia 1.18 [1.15, 1.22] 1.14 [1.09, 1.19] 1.10 [1.07, 1.13] Venous thromboembolism 1.29 [1.24, 1.33] 1.26 [1.21, 1.31] 1.26 [1.21, 1.30] Conduction disorder 1.19 [1.17, 1.21] 1.19 [1.16, 1.21] 1.15 [1.13, 1.17]