Fully automated analysis of cardiotoxicity surveillance echocardiograms: a direct test–retest precision comparison with experts

Serial imaging demands precision—particularly for cardiotoxicity screening where small changes gatekeep major decisions. Automated measurement using artificial intelligence (AI)-based techniques should reduce variability and increase confidence for detection of true change. We directly compared the precision of fully automated AI vs. expert manual analysis of left ventricular function by echocardiography.

Consecutive cancer patients referred for cardiotoxicity monitoring underwent same-day repeat 2D echocardiography [n = 60, 83% female, mean left ventricular ejection fraction (LVEF) 57 ± 7%], with 3D acquisition where feasible (n = 45). All 2D images underwent blinded analysis by a fully automated analysis software and four experts. Quantitative 2D LVEF measurement was feasible in 96% scans by both methods. Mean absolute difference (MAD) between repeat scans was significantly lower for AI than experts for both 2D LVEF {3.6% [confidence interval (CI) 2.8–4.4] vs. 6.6% [CI 5.3–7.9], P = 0.011} and global longitudinal strain [1.5% (CI 1.1–1.9) vs. 2.5% (CI 2.0–3.0), P = 0.006]. In participants with 3DE datasets, 3D LVEF MAD was lower than manual 2D LVEF [3.6% (CI 2.8–4.6) vs. 6.4% (CI 5.0–8.0), P = 0.006] but comparable to AI-derived 2D LVEF [3.5% (CI 2.9–4.2), P = 0.4]. In a subset (n = 48) also with same-day cardiovascular magnetic resonance (CMR) studies, AI 2DE analysis demonstrated stronger agreement with CMR than expert 2DE analysis [LVEF MAD 5.0% (CI 4.1–5.9) vs. 7.8% (CI 6.6–9.0) P = 0.006].

In cancer patients at risk of cardiotoxicity, the precision of 2D echocardiographic AI analysis exceeds that of experts, matching that of 3D LVEF assessment.

NCT06817044

Some cancer treatments may weaken the heart. As such, at-risk patients have repeat heart scans (echocardiograms) over the course of their cancer treatment. Doctors monitor these scans over time and watch closely for any change; even a small decline in heart function can trigger a major decision, such as pausing life-saving cancer therapy. It is therefore vital that these measurements are as consistent or ‘precise’ as possible.

The problem is that measurements taken by hand can vary widely, and ‘precision’ may be low, even when performed by skilled experts. This makes it hard to know whether a change in heart function across follow-up scans is real or not and whether any action is therefore needed. We tested whether artificial intelligence (AI) software could measure heart function more precisely than experts.

Sixty cancer patients each had two echocardiographic heart scans in close succession on the same day. This approach purposefully reduces the impact of other factors (such as differences in the person or equipment performing the heart scan) that might otherwise mean results differ. We then compared how closely the repeat measurements of heart function matched when performed by AI vs. four echocardiography experts. The AI was more precise, giving closer results between the two scans, whilst the experts’ results varied more.

In short, AI can measure heart function more precisely than experts in patients with cancer. This could give doctors greater confidence that any change seen on a scan is genuine, helping them make safer, better-informed decisions.