The data processing workflow for this study is illustrated in Figure 1. We integrated components of PdfFileAnalyzer, an open-source C#-based project, into our processing pipeline to analyze the ECG PDF files [15]. PdfFileAnalyzer is specifically designed to read, parse, and display the internal structure of PDF documents. Through this process, each PDF was decomposed into multiple text-based representations of its constituent objects. Among these, the stream object files contained the critical coordinate information of the ECG waveforms embedded within the vector graphics. By decoding this specific file, coordinate data were extracted and subsequently transformed into time-series representations of time and voltage using predefined scaling formulas.

Detailed examination of the internal PDF structure revealed that ECG waveforms are encoded as vector-based path objects, consisting of sequential coordinate points connected by drawing commands to form continuous line segments (Figure 2). Exploiting this characteristic enabled accurate reconstruction of signals that closely approximate the original ECG waveforms. These vector representations are stored within StreamObj_N.txt files, which contain sequences of Cartesian coordinates associated with PDF drawing commands, such as Move-to (M), Line-to (L), and Stroke (S), defining the geometric paths of each ECG lead. A representative excerpt for lead I, together with the interpretation of the associated commands, is presented in Multimedia Appendix 1.

In parallel with waveform reconstruction, the PDF parsing process also enabled extraction of 8 machine-generated clinical parameters embedded as textual annotations within the ECG report, including ventricular rate; PR interval; QRS duration; QT interval; QTc; and the electrical axes of the P wave, QRS complex, and T wave. These parameters, automatically generated by the ECG acquisition system, were used as the reference standards (ground truth labels) for both the rule-based analyses and the training and evaluation of the deep learning models. The labels were derived from the automated interpretations produced by the MUSE ECG management system rather than from manual adjudications by expert cardiologists. Although not clinically validated by cardiologists, these labels provided a consistent benchmark for evaluating the models’ ability to reproduce the diagnostic outputs of the original ECG acquisition system.

Notably, under certain cardiac rhythm abnormalities—such as atrial fibrillation and junctional rhythm—specific parameters, particularly the PR interval and P-wave axis, are physiologically undefined due to the absence or inconsistency of discernible P waves. In such cases, these parameters are reported as missing values in the system-generated ECG report. This characteristic was explicitly incorporated into the proposed deep learning framework and leveraged during both training and inference, enabling the model to appropriately account for physiologically undefined parameters.

In the PDF coordinate system, the origin (0,0) is located at the bottom-left corner. Each ECG lead includes a 1 mV calibration signal on the left side, which can be identified as a distinct path object (Figure 3). In this study, the calibration signal corresponds to a vertical span of 1000 y-axis units, yielding a voltage scaling factor of 0.001 mV per unit.

The long lead II path object spans 10 seconds and contains 5000 waveform points, corresponding to a sampling rate of 500 Hz. In the PDF, this path object is represented by exactly 25,000 x-axis units, yielding a horizontal scaling factor of 0.0004 seconds per unit, which is used to convert the x-coordinates into time values for the reconstructed waveform.

On the basis of the above analysis, the time and voltage of each point within the path object can be determined using the following formulas:

The variables Xstart and Ystart represent the x-coordinate and y-coordinate of the starting point of each lead’s waveform within the PDF, respectively. The starting points for all leads are summarized in Table 1.

The overall procedure for processing and extracting ECG information from PDF files is summarized in pseudocode form (Textbox 1). The pseudocode outlines the key steps, including PDF parsing, identification of waveform and reference voltage path objects, coordinate extraction, scaling factor estimation, and transformation into time-series signals for each ECG lead. For comparative evaluation, ECG PDF files were converted to SVG format using Inkscape (version 1.4.3), and the resulting vectorized graphical elements were parsed to reconstruct ECG time-series signals for direct comparison with the proposed method.

To evaluate the accuracy of the proposed signal reconstruction process, the ECG time-series data restored from the PDF files were cross-compared with the original ECG signals exported directly from the underlying MUSE system. The assessment focused on the consistency of sampling frequency and differences in voltage amplitude between the 2 datasets. Quantitative evaluation was performed using standard statistical metrics, including mean absolute error (MAE) with corresponding CIs, as well as correlation and agreement analyses.

The rule-based parameter estimation pipeline was developed using the open-source biomedical signal processing library NeuroKit2 [9]. The long lead II was selected as the primary signal source due to its suitability for rhythm analysis. Preprocessing steps included noise reduction and baseline wander removal implemented through NeuroKit2’s standardized filtering procedures. R-peak detection was performed using a multialgorithm strategy incorporating the NeuroKit, Pan-Tompkins, and Nabian methods to enhance detection robustness [9,16,17]. Ventricular rate was calculated from the mean R-R interval when at least 2 R peaks were successfully identified. Wave delineation was achieved using a discrete wavelet transform–based approach to identify key fiducial points, including the onset of the P wave, the onset and offset of the QRS complex, and the end of the T wave. These delineations enabled the computation of the PR interval, QRS duration, and QT interval, all expressed in milliseconds. The QTc was subsequently calculated using Bazett formula. Electrical axis estimation was based on frontal plane vector analysis utilizing signals from leads I and aVF. Representative amplitudes of the P wave, QRS complex, and T wave were determined for each lead, and the corresponding electrical axes were calculated using the arctangent function, yielding angular measurements in degrees.

Overall, this rule-based pipeline provides an interpretable and standardized approach for ECG parameter estimation and serves as a benchmark for comparison with the proposed deep learning models.

Both proposed deep learning models—DualECGFormer and DualResNetECG—adopt a dual-input configuration to process reconstructed ECG signals. Inputs consist of standard 12-lead ECG signals (2.5 s each) and a long lead II signal (10 s). Both models generate 8 independent outputs corresponding to ventricular rate; PR interval; QRS duration; QT interval; QTc; and the electrical axes of the P wave, QRS complex, and T wave. For each parameter, two types of outputs are produced:

ECG waveforms were directly reconstructed from ECG PDF files without additional signal enhancement procedures, such as filtering, denoising, baseline wander correction, or global amplitude normalization. The reconstructed signals were used as direct model inputs to preserve the original waveform morphology encoded within the source PDF files. To ensure consistent input dimensionality across all recordings, standard 12-lead ECG signals (2.5 s per lead) were truncated or zero-padded to 1250 sampling points, whereas long lead II recordings (10 s) were adjusted to 5000 sampling points. Continuous ECG parameter labels were normalized using z score normalization based on the mean and SD estimated exclusively from the training dataset. Normalization statistics were independently computed for each ECG parameter using valid nonmissing samples. Inverse normalization was subsequently applied during inference and evaluation to restore predictions to their original clinical units.

The DualECGFormer model uses a hybrid convolutional neural network (CNN)-Transformer architecture (Figure 4) [18]. Each input branch is initially processed through a ConvFeature module composed of 1D convolutional layers (kernel size 7, padding 3), followed by batch normalization and Gaussian error linear unit activation. Extracted features are projected into sequential embeddings using a PatchEmbed1D module (kernel size=stride=10), producing token sequences of dimension d_model=128.

The embedded sequences are processed by 4 stacked Transformer blocks (depth=4). Each block contains a multihead self-attention mechanism (d_head=4) to model temporal dependencies, a position-wise feed-forward network, layer normalization, residual connections, and positional embeddings. These components collectively encode both local and long-range interactions within the ECG signal.

The DualResNetECG model uses a 1D residual network (ResNet1D) backbone augmented with Squeeze-and-Excitation (SE) modules (Figure 5) [19,20]. Each input branch passes through an independent ResNet1D, with the basic building unit being a 1D residual block (ResBlock1D). SE modules perform global average pooling along the temporal dimension and use 2 fully connected layers (reduction ratio=8) to achieve channel-wise feature recalibration, modulating the relative importance of different feature channels.

Feature representations from the dual branches are fused and passed to task-specific prediction heads. The multitask output design is identical to that of DualECGFormer, providing both value and existence predictions for each of the 8 ECG parameters.

Both DualECGFormer and DualResNetECG are trained using a composite loss function that jointly optimizes regression and classification objectives. The loss function is formulated as a weighted combination of:

Formally, the total loss L is expressed as:

where λ _exist and λ _value are dynamic weighting factors. This approach ensures balanced optimization between estimation of ECG parameter values and prediction of their presence or absence in physiologically undefined cases.

Model training was performed using the Adam optimizer with an initial learning rate of 1×10⁻⁴, a batch size of 256, and a maximum of 100 epochs. Early stopping with a patience of 10 epochs was applied based on validation performance to reduce overfitting. Hyperparameters were empirically selected according to preliminary validation experiments. All experiments were implemented using PyTorch (version 2.7.0; Meta AI) with CUDA (Compute Unified Device Architecture; version 12.8; NVIDIA Corporation) and conducted on a workstation equipped with an NVIDIA RTX 5080 graphics processing unit (GPU). Random seeds were fixed at 42 to improve experimental reproducibility.

The performance of the rule-based approach (NeuroKit2) and the 2 deep learning models (DualECGFormer and DualResNetECG) for estimating 8 ECG clinical parameters is presented in Table 3.

All 3 methods demonstrated good agreement with the reference standard. DualResNetECG achieved the lowest MAE (1.11 bpm, 95% CI 1.03‐1.21) and RMSE (3.48 bpm, 95% CI 2.52‐4.37), along with the highest coefficient of determination (R²=0.97, 95% CI 0.96‐0.99). NeuroKit2 showed comparable MAE (1.18 bpm, 95% CI 1.01-1.35) but a higher RMSE (5.92 bpm, 95% CI 4.91-6.92) and lower R² (0.88, 95% CI 0.84-0.92). DualECGFormer yielded slightly larger errors (MAE 3.87 bpm, 95% CI 3.67‐4.07) with an R² of 0.87 (95% CI 0.84‐0.90).

For the estimation of ECG temporal intervals, both deep learning models outperformed the rule-based approach. DualResNetECG consistently achieved the lowest errors, with MAEs of 7.27 milliseconds (95% CI 6.92‐7.65) for the PR interval, 9.89 milliseconds (95% CI 9.41‐10.42) for the QRS duration, 11.87 milliseconds (95% CI 11.31‐12.45) for the QT interval, and 13.71 milliseconds (95% CI 13.10‐14.27) for the QTc. Corresponding R² values ranged from 0.65 (95% CI 0.61-0.70) to 0.80 (95% CI 0.77-0.83), indicating substantial agreement with the reference measurements.

DualECGFormer also demonstrated improved performance compared with NeuroKit2, with MAEs of 18.30 milliseconds (95% CI 17.76-18.87) for the PR interval, 20.24 milliseconds (95% CI 19.62-20.87) for the QRS duration, 21.11 milliseconds (95% CI 20.38-21.80) for the QT interval, and 23.19 milliseconds (95% CI 22.31-24.12) for the QTc, and R² values between 0.21 (95% CI 0.18-0.23) and 0.56 (95% CI 0.53-0.59). In contrast, the rule-based NeuroKit2 approach exhibited considerably larger errors, with MAEs ranging from 38.03 milliseconds (95% CI 36.92-39.21) to 62.13 milliseconds (95% CI 60.62-63.64) and negative R² values for these parameters, indicating limited agreement with the reference standard.

For the estimation of the electrical axes of the P wave, QRS complex (R axis), and T wave, DualResNetECG again achieved the best performance, with MAEs of 9.33° (95% CI 8.85-9.82), 8.0° (95% CI 7.45-8.59), and 16.31° (95% CI 15.40-17.19), respectively, and R² values of 0.49 (95% CI 0.44-0.55), 0.77 (95% CI 0.72-0.81), and 0.49 (95% CI 0.46-0.52). DualECGFormer demonstrated intermediate performance, while NeuroKit2 showed higher errors and lower or negative R² values, particularly for the P and T axes.

As NeuroKit2 was not able to reliably determine the existence of certain physiologically defined parameters—particularly the PR interval and P-wave axis—the performance of the deep learning models in detecting the presence of these parameters is summarized in Table 4. For PR interval existence, DualResNetECG achieved superior performance, with an accuracy of 0.96, precision of 0.98, recall of 0.98, specificity of 0.81, and an F₁-score of 0.98. DualECGFormer also demonstrated strong performance, with an accuracy of 0.93 and an F₁-score of 0.96, although with lower specificity (0.60).

Similarly, for P-wave axis existence, DualResNetECG outperformed DualECGFormer, achieving an accuracy of 0.95, precision of 0.98, recall of 0.97, specificity of 0.82, and an F₁-score of 0.97. DualECGFormer yielded slightly lower performance, with an accuracy of 0.92 and specificity of 0.63.

DualResNetECG demonstrated higher specificity and negative predictive value (NPV) than DualECGFormer for both PR interval and P-wave axis existence, reflecting improved identification of physiologically undefined parameters. Specificity increased from 0.60 to 0.81 for the PR interval and from 0.63 to 0.82 for the P-wave axis, while NPV improved from 0.69 to 0.79 and 0.68 to 0.79, respectively.

Consistent patterns were observed in receiver operating characteristic and precision-recall analyses (Figure 7). For PR interval existence, DualResNetECG achieved an area under the receiver operating characteristic curve (AUROC) of 0.978 and an area under the precision-recall curve (AUPRC) of 0.995, compared with an AUROC of 0.939 and an AUPRC of 0.990 for DualECGFormer. For P-wave axis existence, DualResNetECG achieved an AUROC of 0.977 and an AUPRC of 0.995, whereas DualECGFormer achieved an AUROC of 0.941 and an AUPRC of 0.990.