A deep learning framework trained on 11,330 expert-annotated ECG leads from 1,030 UK Biobank participants achieved mean absolute errors of just 7.7 ms (PR), 7.5 ms (QRS), and 4.9 ms (QT) interval measurements — outperforming CardioSoft, an open-source signal-processing toolbox, and a wavelet-based delineation method. Critically, prolonged corrected QT intervals detected by the deep learning model predicted major adverse cardiovascular events (MACE) with a hazard ratio of 2.9 (95% CI 2.1–4.0) across 46,749 participants over a median 4-year follow-up, compared to hazard ratios of 1.7 and 1.1 for wavelet and CardioSoft methods respectively — suggesting measurement precision meaningfully amplifies clinical signal.
This finding matters because QTc prolongation is already an established cardiovascular risk marker linked to arrhythmia and sudden cardiac death, but automated measurement inaccuracies have historically diluted its predictive power in large-cohort studies. The near-tripling of MACE hazard ratio compared to the standard CardioSoft tool suggests prior population-level studies may have substantially underestimated QTc's true prognostic value. For adults undergoing routine ECGs, more accurate AI-derived intervals could sharpen risk stratification without additional testing. Limitations include a relatively short 4-year follow-up, the predominantly European UK Biobank ancestry limiting generalisability, and reliance on a single annotation cohort for training. As a preprint not yet peer-reviewed, these hazard ratio estimates and validation metrics require independent replication before clinical adoption. If confirmed, this framework represents a meaningful methodological advance rather than merely an incremental refinement.
