Improved measurement precision for cancer cardiotoxicity screening using fully automated analysis

Abstract

Listen

Translate article

Abstract Background Serial surveillance screening for cardiotoxicity demands measurement precision to detect dysfunction early and avoid inappropriate treatment interruptions. 3D LVEF and GLS imaging are therefore recommended, or CMR where image quality is suboptimal [1]. Automated (AI) analysis of CMR has demonstrated ‘superhuman’ precision [2], however the impact of AI on echocardiographic precision is unknown. Purpose To compare human versus AI analysis for LVEF and GLS using echocardiography in oncology patients at risk of cardiotoxicity. Methods Adult oncology patients underwent same day repeat echocardiography (GE Vivid9 machine) using recommended protocols (ethics 16/LO/1815). Manual analysis by a single blinded expert observer was compared to AI analysis (of 2D images) using two commercial AI packages (AI1, AI2) and a third AI package currently undergoing regulatory approval (AI3). A subset of patients also underwent paired same day CMR, analysed using commercial software (Circle CVI). Primary outcome measure was mean absolute difference (MAD) between repeat scans, and within subject co-efficient of variation (WSCoV), minimal detectable difference (MDD) and Bland-Altman limits of agreement were also analysed. Statistical analysis was performed in R, with n=10,000 bootstrap to estimate confidence intervals. MAD was compared with a Mann-Whitney U test. Results 61 cancer patients (median age 51, 84% female) underwent same day repeat echocardiography. Image quality was graded acceptable for LVEF analysis by humans in 92% cases for 2D and 74% of 3D images. AI analysis was feasible in 96% (AI1), 93% (AI2) and 100% (AI3) of adequate 2D studies. Median LVEF across the group was 58% (IQR 50 – 66%). MAD for 2D LVEF (n=56) was similar between humans and AI software packages AI1, AI2 and AI3 respectively (4.2% [3.5-49] vs 4.8% [3.7–5.9], 5.2% [4.2-6.4], 3.5% [2.7-4.2]), Figure 1 and Table 1. However for LVEDV and LVESV, AI3 had significantly lower MAD (6.8ml [5.5-8.1] vs 11.8ml [9.5-14.2] and 3.4ml [2.7-4.1] vs 6.5ml [5.3-7.7], p<0.001) compared with humans. MDD for EDV was 10.5ml [8.2-12.3] for AI3 vs 20ml [16.1-23.4] for human 2D LVEF (p<0.05). For GLS, MAD was similar between humans, AI2 and AI3 though significantly higher for AI1 (2.6% [2.2-3.0] vs 1.4% [1.1-1.7], p<0.001). For participants (48) with matched CMR paired data, MAD for CMR LVEF was significantly lower than human, AI1 and AI2 for LVEF, LVEDV and LVESV respectively [p<0.05], but not significantly different to AI3 for LVEF (3.9% [3.2-4.7] vs CMR 2.9% [2.3-3.4] or 3D 3.3% [2.7-3.9] p>0.05). Conclusion In oncology patients at risk of cardiotoxicity, newer fully-automated AI based echocardiography analysis improves measurement precision, with similar performance to 3D echocardiography and approaching CMR. Feasible even with challenging images, this should improve both confidence and efficiency for cardiotoxicity surveillance in cancer patients.