QTcNet: a deep learning model for direct heart rate corrected QT interval estimation

Figures

• 

  Figure 1. Comparison of algorithm-based QTc estimates (machine measurements), QTcNet predictions, and expert annotations on the QTcMS and EDMS datasets. Panel A shows QTcNet predictions (x-axis) vs. machine measurements (y-axis) on EDMS; panel D shows machine measurements (x-axis) vs. expert annotations (y-axis) on QTcMS; and panel G presents QTcNet predictions (x-axis) vs. expert annotations (y-axis). Each point corresponds to an individual ECG, with the diagonal line marking perfect agreement. Six outlier ECGs are highlighted in blue, black, purple, red, yellow, and green on the scatter plots. Their corresponding waveforms (lead II) are displayed in panels B, C, E, F, H, and I, each with borders in the matching colour. Outliers were reduced with QTcNet despite high levels of noise and artefacts (panel G).

• 

  Figure 2. Direct comparison of QTcNet performance compared to machine measurements in QTcMS. Panel A shows the mean absolute error of QTcNet predictions and machine measurements vs. human expert annotations across varying heart rates. Error bars indicate 95% confidence intervals. Stars denote the outcome of the appropriate two-sided paired test applied to each interval (paired Student’s t-test when the paired differences passed the Anderson–Darling normality test, otherwise Wilcoxon signed-rank): *P < 0.05, **P < 0.01, ***P < 0.001. Panel B compares the spread of error of QTcNet (blue) and machine measurements (orange) vs. expert annotations, illustrating both a systematic overestimation and a wider spread of error with machine measurements compared to QTcNet. Bland–Altman plots illustrate pairwise agreement of QTcNet (panel C) and machine measurements (panel D) vs. expert annotations. Horizontal dashed lines indicate upper and lower limits of agreement.

• 

  Figure 3. Subgroup performance of QTcNet (panel A) and machine measurements (panel B) in the QTcMS dataset. Mean absolute error (MAE) and root mean squared error (RMSE) is given for each subgroup as the absolute difference compared to the overall cohort. SR, sinus rhythm; PVC, premature ventricular complexes; AFib/AFlu, atrial fibrillation or atrial flutter; BBB, bundle branch block.

• 

  Figure 4. Comparison of model performance as a function of training set size (log-scale). Panel A shows the mean absolute error (MAE) ratio and panel B shows the root mean square error (RMSE) ratio relative to a start sample size of n = 10 000 evaluated on five different sets: MIMIC (blue), EDMS (orange), QTcMS Expert (green), PTB (red), and ECGRDVQ (purple). Expert-based datasets where marked as bold line. The vertical dashed line highlights the last point of balanced EDMS and MIMIC training data.

• 

  Figure 5. Integrated Gradients (IG) maps illustrating feature relevance for QTc prediction. Panel A shows dataset-wide aggregated importance scores computed using IG for each ECG lead and aligned to median beats. Black lines represent the average ECG signals across the test set, while the red overlay highlights regions the model identified as relevant for QTc prediction. Darker areas indicate higher attribution values. Panels B–D display three individual test set ECGs alongside their respective importance scores, illustrating how the model’s focus can vary across different signal morphologies.