PO-04-209 A NOVEL MACHINE LEARNING (ML) ALGORITHM TO MEASURE THE CORRECTED QT INTERVAL FROM MOBILE ECGS WITH HIGH ACCURACY AND PRECISION

Abstract

The heart rate-corrected QT interval (QTc) is an important feature of the electrocardiogram (ECG), as it is an important prognosticator for ventricular arrhythmia and sudden cardiac death. The 12-lead (12L) ECG is used most often for assessment of QTc, and algorithms that have extended QTc detection to mobile ECGs with fewer than 12 leads are prone to wide sample size deviation. For clinical applications requiring serial QTc monitoring, such as titration of QTc-prolonging drugs, both accuracy and precision are critical. The aim was to develop a novel machine-learning (ML) algorithm that measures QTc from mobile ECG data formats with high accuracy as well as precision. The acceptance criteria was defined as error and standard deviation within 20 milliseconds (ms), the human interobserver variability observed for manual ECG measurements (gold standard). The ML algorithm was trained using a database of 13,000 cardiologist-annotated 12L ECG heartbeats, including the following steps: 1) pre-processing including resampling to 300 Hz, 2) normalization of amplitudes using min-max scaling 3) noisy beats with excessive baseline wander were discarded, and 4) applying TabNet and XGBoost models for detection of the end of T-Wave. This ML model for detection of the T-wave endpoint was merged with a signal processing method for detection of the Q-point. The algorithm was then validated on a subset of 1,239 cardiologist-annotated heartbeats, sufficient to assess strong agreement between software and manual methods (95% CI ± 0.2 std, Bland-Altman approach). Results showed high AUC, sensitivity, and specificity for detecting QTc above clinically meaningful thresholds. Mean QTc error was 5.24±7.22 ms, meeting predefined acceptance criteria. The ML algorithm was then applied to a dataset of 150 6L mECGs (AliveCor; Mountain View, CA), which comprise an all-comer population of a combination of healthy (∼40%), outpatient clinic (30%), and hospitalized patients (30%), with error 13.7±7.78 ms, again meeting acceptance criteria. Algorithm validation surpassed the defined acceptance criteria and demonstrated that the algorithm can provide and potentially surpass human QTc measurement capabilities. By enabling automated, dynamic QTc analysis, the algorithm offered the precision required for outpatient use.