Uncertainty Calibrated Deep Regression for QT Interval Measurement in Reduced Lead Set ECGs

Accurate measurement of the QT interval on the electrocardiogram (ECG) is important in assessing a patient's cardiac safety and predisposition to potentially lethal arrhythmias. However, many ECG monitoring systems provide measurements of the QT interval without a level of confidence or certainty in the measurement. To address this, we present an uncertainty aware deep learning model for the measurement of QT intervals on reduced-lead set ECGs. Our model incorporates both aleatoric and epistemic uncertainty prediction, thereby presenting a more comprehensive picture. The model employs an InceptionTime architecture to ensure precise predictions. A novel prediction head captures both types of uncertainty, allowing the model to quantify its inherent variability. Calibration ensures the uncertainties align with true errors. We evaluate the model on a large dataset with 12-lead, 6-lead, and 1-lead ECGs. The model shows excellent agreement with reference values, with mean differences less than 9ms across all lead sets. Calibration plots reveal well-calibrated uncertainties for all models, with an expected normalized calibration error less than 20%. This work demonstrates the potential of uncertainty-aware deep learning for QT interval prediction in reduced-lead ECGs. It offers a valuable tool for informed decision-making by quantifying the inherent variability in each prediction.