Online Poster Portal

Background: Patients with cancer may experience chemotherapy-related cardiac dysfunction (CTRCD) as a result of treatment. Left ventricular ejection fraction (LVEF) is used to screen for CTRCD; however, it has high interobserver variability. Artificial intelligence (AI) applied to echocardiography can potentially detect CTRCD earlier and more reliably.

Methods: We followed a cohort of 6,052 cancer patients undergoing treatment at Kaiser Permanente Northern California who had a pre-therapy echocardiogram within one year of initiation with a normal LVEF ≥ 53% and a post-therapy echocardiogram within 2 years. CTRCD was defined using guidelines from the American Society of Echocardiography. Patients were classified as having CTRCD when evaluated by clinical LVEF; CTRCD when evaluated by AI LVEF; patients with low baseline LVEF < 53% who were not recognized clinically; and no CTRCD by either AI or clinical assessment of LVEF. To better understand consistency in clinical assessments for CTRCD, we asked 3 cardiologists to re-evaluate LVEF in 5 patients who had clinical CTRCD but not AI CTRCD and in 12 patients with AI CTRCD but not clinical CTRCD. Cardiologists were blinded to all identifying information, including the paired nature of the studies.

Results: Out of the 6,052 patients, 528 (6.7%) were diagnosed clinically with CTRCD. A higher CTRCD incidence was found in patients undergoing treatment with immune checkpoint inhibitors in 69 (10.9%) patients, tyrosine kinase inhibitors in 67 patients (10.7%), and anthracyclines in 273 (10.5%) patients. On blinded re-read, cardiologists confirmed CTRCD in only 1 of the 5 patients previously identified with clinical CTRCD and confirmed 8 out of 12 AI CTRCD patients as abnormal. Our findings suggest that clinicians are neither sensitive nor specific for detecting CTRCD. On AI re-analysis of surveillance echocardiograms, an additional 336 (6.0%) patients were found to have CTRCD and 281 (4.6%) patients were found to have reduced LVEF at baseline. Both groups had a significantly increased proportion of patients with elevated brain natriuretic peptide compared to those without CTRCD (proportions z-test p < 0.001). Kaplan-Meier curves for mortality and hospitalization were similar for patients with recognized CTRCD and high-risk patients identified by AI (log rank p = 0.17 for mortality; log rank p = 0.06 for hospitalization). All three subsets had significantly increased event rates of mortality, hospitalization, emergency department visits, and heart failure exacerbation from those without CTRCD (log rank p < 0.001).

Conclusion: Clinical evaluation for cardiotoxicity is inconsistent. AI identifies additional patients with similar risk profiles to patients with clinical CTRCD. Our work suggests AI can serve as an additive tool for surveillance and identify patients at risk for cardiotoxicity who may otherwise be missed.