Detection of Cancer Therapy-Related Cardiac Dysfunction Using ECG

Administration of anticancer drugs in cancer treatment can cause side effects on the heart. If such cancer therapy-related cardiac dysfunction (CTRCD) can be detected early, adjusting the treatment plan may prevent it from becoming severe. However, capturing the subtle changes that appear before a decline in cardiac function becomes evident is not straightforward. In this research, we are developing systems that detect signs of CTRCD from 12-lead electrocardiograms obtained in routine examinations. A 12-lead ECG is a multivariate time series that observes the heart from different directions, so information is contained both in local changes within waveforms and in relationships among leads. We are therefore investigating models that integrate features across multiple time scales, as well as methods that explicitly model inter-lead relationships as a graph structure. To address the limited availability of labeled data that is characteristic of medical datasets, we are also exploring self-supervised representation learning. Technical approaches: - Multi-scale feature learning integrating CNN, RNN, and attention mechanisms - Modeling of inter-lead relationships using graph neural networks - Self-supervised representation learning for limited labeled data We are advancing verification using real clinical data through joint research with the National Cancer Center Japan.