Performance of fully automated AI-based left ventricular strain measurement in cancer patients treated with immune checkpoint inhibitors

Background: Immune checkpoint inhibitors (ICIs) have improved clinical outcomes for cancer, however, their use is associated with cardiotoxicity. While left ventricular global longitudinal strain (LV-GLS) is considered the standard of care for early myocardial damage, the routine use of LV-GLS is limited by its time-consuming nature, and the need for expert cardiologists. AI-based LV-GLS applications may offer a cost-effective alternative.
Objective: To compare standard, semi-automated LV-GLS analysis to a fully-automated AI-based LV-GLS software, and to evaluate its performance in predicting cardiovascular (CV) events in patients treated with ICIs.
Methods: We retrospectively analyzed consecutive cancer patients who underwent baseline transthoracic echocardiography before the initiation ICIs. LV-GLS analysis was conducted using both standard semi-automated, vendor-independent software (TomTec) and fully-automated AI-based software (Us2.AI). CV events were defined as a composite of myocarditis, acute coronary syndrome, heart failure and arrhythmia. We examined the correlation between the two techniques and their ability to predict future CV events using logistic regression and receiver operating characteristic (ROC) curves.

Results: Out of 183 patients (mean age 65±12 years, 43% females), 157 (86%) had adequate image quality for strain analysis. During a median follow up period of 48 months (IQR: 34-61), 23 (13%) patients experienced CV events. A significant difference was observed between the baseline mean standard LV-GLS and AI-based LV-GLS (18±2.9 vs. 20±2.9, P<0.001), with a moderate positive correlation (r=0.38, P<0.001). Lower standard LV-GLS was predictive of CV events (OR = 0.83, 95% CI 0.71-0.98, P=0.03), whereas AI-based LV-GLS did not (OR = 0.96, 95% CI 0.82-1.14, P=0.686). In ROC analysis, standard LV-GLS showed superior area under the curve compared to AI-based measurements, but the difference was not statistically significant (AUC 0.62 vs. 0.53, P=0.275).
Conclusions: Fully automated AI-based LV-GLS provided higher values compared to standard, semi-automated LV-GLS, with a moderate correlation. Standard LV-GLS was predictive of future CV events.