Direct Reconstruction of High-Fidelity Electrocardiogram Signals From Vector-Based PDF Files With Integrated Deep...

Electrocardiograms (ECGs) are commonly stored in PDF, particularly as vector-based files generated by ECG management systems. Previous studies have demonstrated that ECG signals can be extracted through PDF-to-Scalable Vector Graphics (SVG) conversion, highlighting the potential to reconstruct waveform signals from vector graphics. These reconstructed signals further enable the derivation and prediction of clinically relevant ECG parameters.

This study aimed to develop an integrated framework for direct reconstruction of high-fidelity ECG signals from vector-based PDF files and for simultaneous estimation of multiple clinically relevant ECG parameters using deep learning.

In this retrospective methodological study, 50,000 twelve-lead ECG PDFs generated by a MUSE system (2015-2024) were analyzed. A direct PDF parsing pipeline was developed to extract vector path objects and reconstruct time-series signals without intermediate format conversion. Reconstruction accuracy was evaluated against the original system-exported signals. Two deep learning models, DualECGFormer and DualResNetECG, were developed to estimate 8 specific ECG parameters. The reference standards for these parameters-including ventricular rate; PR interval; QRS duration; QT interval; corrected QT interval (QTc); and the electrical axes of the P wave, QRS complex, and T wave-were derived from machine-generated values within the MUSE system database. Performance was compared with a rule-based approach (NeuroKit2).

The proposed method achieved high reconstruction fidelity, with mean absolute errors (MAEs) below 1×10-3 mV across all leads. Compared with an SVG-based workflow, the direct parsing approach reduced processing time by approximately 4.6-fold. For parameter estimation, deep learning models outperformed the rule-based method for most parameters. DualResNetECG achieved the best overall performance, with MAEs of 1.11 bpm for ventricular rate, 7.27 milliseconds for PR interval, 9.89 milliseconds for QRS duration, 11.87 milliseconds for QT, and 13.71 milliseconds for QTc. For electrical axis estimation, MAEs ranged from 8.0° to 16.31°. The model also demonstrated reliable detection of physiologically undefined parameters (PR interval and P-wave axis), achieving an area under the receiver operating characteristic curve of up to 0.978.

This study presents an efficient and scalable framework for direct extraction of ECG signals from MUSE-generated vector-based PDFs and integrated multiparameter estimation using deep learning. The approach achieves high reconstruction accuracy and competitive predictive performance, supporting its potential utility for large-scale retrospective MUSE ECG analysis.