EventsThe 13th International Electronic Conference on Sensors and Applications
Published
This submission belongs to the session S9. Student Session of the event The 13th International Electronic Conference on Sensors and Applications
Published date
16 Nov, 2026
Academic Editor
author-avatarFrancisco Falcone
Citation
Kaveetha Herath, Lakshan Madhushanka, Digitizing Legacy ECG Sensor Outputs for Automated Multi-Class Cardiac Abnormality Classification Using Deep Learning, in Proceedings of The 13th International Electronic Conference on Sensors and Applications, 18 November–20 November 2026, MDPI: Basel, Switzerland
Share
Email
Facebook
Twitter
LinkedIn

Digitizing Legacy ECG Sensor Outputs for Automated Multi-Class Cardiac Abnormality Classification Using Deep Learning

1. Department of Computer Engineering, Faculty of Engineering, University of Sri Jayewardenepura, No 120, Kekunagahawatta, Polgasowita Road, Mattegoda, Sri Lanka
Abstract

Electrocardiogram (ECG) is one of the most widely practiced biomedical sensing modalities for cardiac assessment to investigate hear related conditions. In many healthcare systems, particularly across South Asia and other regions with finite resources, ECG sensor outputs are typically stored as paper prints, scanned images, or photographic records rather than machine-readable digital signals. This makes large parts of clinically valuable sensor data unavailable to modern AI-based diagnostic systems which rely on structured time-series input.

This study presents a framework for recovering ECG waveforms from paper based ECG records which also allows automated multi-class cardiac classification. The digitization pipeline uses classical image processing: Hough transform-based isoelectric baseline localization for colour-grid ECGs, argmin-trace detection for grayscale inputs, colour-channel grid removal, and median-pixel waveform extraction with calibrated millivolt and millisecond output. Recovered signals from individual lead images are assembled into a unified 12-lead tensor, with missing leads handled by dynamic zero-padding. A 1D ResNet classifier (ECGResNet1D) is trained on 21481 PTB-XL records across five diagnostic superclasses: Normal, Myocardial Infarction, ST/T-wave Changes, Conduction Disturbances, and Hypertrophy, using weighted sampling, label smoothing, and signal augmentation to address class imbalance and digitization-induced noise.

Digitization fidelity assessment with respect to the PTB-XL ground-truth waveforms reveals, according to preliminary results, near-perfect morphological correlation for clean paper inputs (Pearson r = 0.966, SNR=12.09 dB) and acceptable recovery for red-grid paper (r=0.661, SNR=2.21 dB). The classifier reaches 72.0% validation accuracy with a macro-averaged F1 of 0.618. A systematic preliminary lead degradation experiment shows graceful performance degradation from macro F1=0.116 with one rhythm strip to 0.598 with all 12 leads.

The proposed framework demonstrates that a rule-based image processing approach, without any trained digitization model, is sufficient to extract usable cardiac signals from old paper ECG records and support automated diagnosis in clinical settings where digital records do not exist.

Keywords
ECG digitization
legacy sensor data recovery
cardiac abnormality classification
1D residual network
biomedical sensing
PTB-XL
Manuscript
[Sciforum-194239] - [Manuscript PDF] - [Oral contribution - file uploaded].pdf
Investigation of Features of Three Wearable PPG Multi-Sensors for Measurement in Low Magnetic Field MRI Tomograph
An Energy- and Area-Efficient Modified Blowfish Encryption Scheme for Secure Healthcare IoT Sensor Nodes