The recognition and utilization of ECG signals from clinical reports as PDF files

Electrocardiograms (ECGs) are routinely stored as static PDF reports in clinical archives, limiting their direct use for computational analysis, heart rate variability (HRV) assessment, or retrospective research. This thesis develops and validates a comprehensive methodological pipeline for the recognition, digitization, and analysis of 12-lead ECGs extracted from clinical PDF documents. The workflow integrates automated lead cropping, image-based waveform reconstruction, structured text extraction, physiological calibration, and a custom R-peak correction algorithm designed to ensure cross-lead temporal consistency is an essential requirement for reliable HRV computation in digitized signals. The methodology was applied to resting ECGs from heart failure patients enrolled in a six-month supervised exercise rehabilitation program, compared with a control cohort that did not participate in structured physical activity. The control group received no exercise recommendations and continued their routine daily activities under standard clinical care. Time-domain HRV indices were computed directly from the digitized ECG waveforms, whereas conventional ECG intervals (RR, PP, PR, QRS, QT/QTc, axes) were extracted from the automated clinical information included in the PDF report header rather than recalculated from the reconstructed signal. Statistical evaluation included paired t-tests or Wilcoxon signed-rank tests, correlation analyses with age, and two-way ANOVA assessing the effects of Intervention and Age Category on Δ values (AFTER − BEFORE). Results showed modest temporal changes in the control group, including significant prolongation of RR, PP, and QT intervals and a reduction in ventricular rate, reflecting mild physiological deceleration over time. In contrast, the exercise group exhibited stable interval durations, with only a shift in the R-wave axis reaching significance. Two-way ANOVA confirmed significant Intervention effects for RR, PP, QT intervals and ventricular rate, suggesting that exercise helped preserve electrophysiological stability relative to the natural slowing observed in controls. However, no significant improvements in HRV indices were detected in the exercise group, a finding consistent with expert clinical insight that HRV measures derived from ultra-short (10 second) ECGs have limited sensitivity for capturing chronic autonomic adaptations. Overall, this work demonstrates the technical feasibility and clinical potential of extracting physiologically meaningful information from PDF-based ECGs. The proposed pipeline provides a reproducible framework for transforming static ECG archives into analyzable digital biosignals, supporting retrospective studies, automated analytics, and integration into machine-learning pipelines. Beyond methodological contributions, the study underscores both the opportunities and limitations of short-segment digitized ECGs for HRV analysis, highlighting the need for longer recordings in future research to more accurately quantify autonomic responses to exercise rehabilitation.