EchoRisk: A Multicentre Echocardiography Dataset and Benchmark for Cardio-Oncology

Therapy-induced cardiotoxicity is a major complication for breast cancer patients, often leading to treatment interruptions and long-term cardiovascular morbidity. While echocardiography is the standard for monitoring, manual interpretation is prone to observer variability and often detects cardiac injury only after significant damage has occurred. This paper introduces EchoRisk, a curated, multicenter, longitudinal echocardiography dataset designed to facilitate the development of automated AI tools for cardio-oncology.

EchoRisk is derived from the CARDIOCARE prospective study and includes 422 patients across five European sites. The dataset features 2,159 echocardiography videos collected at up to five longitudinal timepoints. The authors define three distinct tasks:

• Task 1 (LVEF Estimation): Automated estimation of left ventricular ejection fraction from cine video.
• Task 2 (LV Dysfunction Classification): Detecting left ventricular dysfunction based on clinical guidelines.
• Task 3 (Early Cardiotoxicity Prediction): Forecasting future cardiotoxicity using only pre-therapy baseline echocardiography.

The researchers established baseline performance using an R(2+1)D video backbone with LSTM aggregation. The models performed well on Tasks 1 and 2, achieving a mean absolute error of 4.98 percentage points for LVEF estimation and an AUC of 0.849 for LV dysfunction classification. However, Task 3 proved difficult; the best-performing video-based model achieved an AUC of 0.541, which is statistically indistinguishable from a clinical reference model based on age and baseline LVEF. This indicates that early prediction of cardiotoxicity from baseline imaging alone is an open problem that current end-to-end architectures have yet to solve.

By providing a standardized, multicenter benchmark, EchoRisk enables researchers to move beyond single-site studies and develop more robust, reproducible AI models for cardio-oncology. The dataset highlights the gap between current clinical capabilities and the potential for early, non-invasive risk stratification, providing a clear target for future algorithmic innovation in oncology patient care.