A new open dataset wants to make it easier for AI to catch heart damage before it forces cancer patients off treatment.
Researchers released EchoRisk, a multicentre longitudinal echocardiography dataset built from 422 breast cancer patients enrolled in the EU-funded CARDIOCARE study across five European sites. The collection includes 2,159 echocardiography videos spanning 1,123 clinical exams, each tagged with cardiotoxicity labels — a combination that, according to the paper, has not existed in a curated public form before. The dataset anchors three benchmark tasks: estimating left ventricular ejection fraction from video, classifying LV dysfunction across timepoints, and — the hardest one — predicting cardiotoxicity from a single pre-therapy scan before any drug has been administered.
Therapy-induced cardiotoxicity is the top non-oncological reason breast cancer treatment gets interrupted, so catching it early has real clinical stakes. The researchers established a baseline using an R(2+1)D video backbone with LSTM aggregation pretrained on Kinetics-400; it handles cardiac function assessment reasonably well, but early prediction from a single baseline video remains, by their own description, an open problem.
The timing is deliberate — EchoRisk is the primary reference dataset for the EchoRisk-MICCAI 2026 challenge, which should attract competitive modeling from the medical-imaging community. Whether a public benchmark actually closes the gap on early prediction, or just formalizes how hard the problem is, will depend on what the challenge submissions produce.