Utilising artificial intelligence on serial electrocardiogram recordings for personalised clinical risk prediction

Clinical risk prediction models are increasingly used in healthcare to guide clinical decision-making. Artificial intelligence-enhanced electrocardiogram (AI-ECG) approaches have shown promise in risk prediction. However, to date, AI-ECG models face a critical limitation: they only use a single ECG as input, neglecting information present in previous and future ECGs, despite their well-recognised clinical utility. Capitalising on recent advancements in longitudinal medical AI, we propose the longitudinal AI-ECG risk estimator (LAIRE) - the first purely deep learning-based platform to harness serial ECGs for improved predictions. Prototypes use a bespoke AI architecture using sequences of pretrained ECG representations and leverage the power of transformers, which are state-of-the-art in sequence modelling. Preliminary testing on ~50,000 patient ECG sequences from the Beth Israel Deaconess Medical Center dataset, suggest significant improvements over single-ECG models predicting time-to-mortality. We will develop the LAIRE platform to forecast death and future clinical risks, including sudden cardiac death, myocardial infarction, and heart failure. Interpretability analyses will enhance clinical insights into longitudinal ECG changes, while the added incorporation of image-based ECG recordings and multimodal data may improve application, performance and generalisability. If successful, LAIRE will represent a key milestone in clinical risk prediction, serving as an important tool for clinical decision-making.