AI Research Portal

This paper introduces ECGLight, an end-to-end framework designed for the compute-light digitization of paper ECGs and subsequent screening for Myocardial Infarction (MI). The system reconstructs high-fidelity 12-lead ECG signals from scanned images and then uses lightweight classifiers to detect MI, differentiate pre- and post-procedural states, and identify Occlusion Myocardial Infarction (OMI). A key empirical claim is that the system can perform this entire process on CPU-only resources in under 30 seconds per ECG, achieving high accuracy on benchmark datasets.

The paper addresses several bottlenecks in utilizing paper ECGs for modern AI-driven diagnostics:

Limited Connectivity and Computational Capacity: Many remote clinics rely on paper ECG printouts due to insufficient internet connectivity or computational power, preventing access by AI-based decision support systems.
High Resource Requirements of AI Models: Existing AI models for ECG analysis often demand significant computational resources and strong internet connections, making them unsuitable for low-resource environments.
Overlooked Conditions and Delayed Therapy: The inability to access and analyze paper ECGs promptly can lead to missed diagnoses of critical conditions like acute coronary occlusion (ACS), delaying essential reperfusion therapy.
Disparate Digitization and Classification: Prior work has typically tackled ECG digitization and diagnostic classification as separate problems, lacking an integrated, efficient, and on-device solution.

An end-to-end, compute-light, on-device framework for paper ECG digitization and MI screening.
A lightweight segmentation module based on YOLOv11 for waveform extraction from noisy paper ECGs.
Automated lead identification, layout detection, and reference-pulse detection for physically grounded amplitude scaling.
A pipeline that reconstructs calibrated 12-lead ECG time-series signals from paper images.
Integration of downstream classification for multiple clinically relevant endpoints (MI vs. Normal, Pre- vs. Post-procedural MI, OMI vs. Non-OMI).
Use of SHAP for physiologically grounded, interpretable attributions to support clinical understanding.
Demonstration of the system's ability to run on CPU-only resources with low latency (<30s per ECG).

The ECGLight framework operates in several stages:

Input: A paper ECG image is captured via smartphone or scanner.
Paper ECG Digitization (YOLOv11-Based):
- Waveform Segmentation: YOLOv11 is used for instance segmentation to identify and segment individual ECG leads.
- Lead Identification & Layout Detection: The system identifies the 12 leads (I-III, aVR-aVL, V1-V6) and their spatial arrangement on the page.
- Reference Pulse Detection & Scaling: A calibration pulse (typically 1 mV, 0.2 s) is detected to determine pixel-to-mV and pixel-to-time scaling factors.
- Waveform Reconstruction: Using the identified leads, layout, and scaling factors, physically calibrated 12-lead digital ECG time-series signals are generated. Post-processing includes baseline trimming, per-lead normalization, and R-peak detection.
Lightweight Classification (Time-Series Classifiers): The reconstructed digital ECG signals are fed into lightweight, edge-optimized classifiers for tasks such as MI vs. Normal, Pre- vs. Post-procedural ECG, and OMI vs. Non-OMI. These classifiers are designed for fast inference on CPU-only devices.
Interpretability (SHAP): Shapley Additive Explanations (SHAP) are used to provide insights into the classification decisions. This includes identifying important leads and specific regions within the heartbeat (e.g., QRS, ST-T segments) that contribute most to the prediction, enhancing transparency and clinical understanding.
Output for Clinicians: The system provides diagnostic predictions along with an interpretability report, empowering doctors in low-resource settings.

Architecture: The core digitization component utilizes YOLOv11 for segmentation, lead identification, and reference pulse detection. Classification is performed by lightweight time-series models, and interpretability is achieved using MLP and SHAP.
Key Equations:
- Intersection over Union (IoU): IoU = |A ∩ B| / |A ∪ B| (Equation 1)
- Dice Coefficient: Dice = 2 × |A ∩ B| / (|A| + |B|) (Equation 2)
- Pearson correlation (r): r = Σ((xi - x̄)(yi - ȳ)) / (√Σ(xi - x̄)² √Σ(yi - ȳ)²) (Equation 3)
- Root Mean Square Error (RMSE): RMSE = √(1/n Σ(Yi - Ŷi)²) (Equation 4)
- Signal-to-Noise Ratio (SNR): SNR = 10 log10 (Σŷi² / Σ(yi - ŷi)²) (Equation 5)
Loss/Objective: Not explicitly detailed for the YOLOv11 model, but standard object detection losses are implied. Classification models use standard supervised learning objectives.
Training:
- YOLOv11 models were fine-tuned for 100 epochs using AdamW optimizer, cosine learning rate schedule, and FP16 acceleration.
- For segmentation, both "Full-Image" and "Patched" regimes were used. The patched regime involved overlapping tiles (e.g., 1280x1280).
- Classification models were trained for 200 epochs with batch size 64, using a 75/25 train-test split. Kernel-based models used 20,000 estimators.
- MLP for SHAP analysis used one hidden layer (100 neurons), ReLU, Adam, and was trained for 200 epochs with early stopping.
Augmentation: Synthetic augmentations included illumination variation, Gaussian noise, motion blur, rotation, handwritten annotations, paper creases, and faded grid lines.
Data Handling:
- Training Data: PTB-XL dataset (21,799 ECGs, 18,869 patients) was used, with synthetic renderings generated by ECG-Image-Kit. Patched training used ~323,000 patches from 11,000 recordings.
- Validation/Testing Data: PTB-XL evaluation subset (1,600 ECGs, 1,574 patients) and ECG-Matrix dataset (hospital-acquired ECGs) were used.
Implementation Details: The system runs on CPU-only resources. Inference time for the end-to-end pipeline is approximately 25-30 seconds per ECG on an Intel Core i9-13900H laptop.

Datasets: PTB-XL (synthetic and real paper renderings), ECG-Matrix (real hospital paper ECGs).
Metrics:
- Digitization: IoU, Dice, Pearson correlation (r), RMSE, SNR.
- Classification: Accuracy, Precision, Recall, Specificity, F1-Score, AUROC, AUPRC.
Baselines: Comparisons were made against leading challenge submissions like SignalSavants, BAPORLab, USST_Med, Ahus AI Lab, and wavie_ABI for digitization. For classification, various deep learning (InceptionTime, LSTM-FCN, MCDCNN, ResNet, CNN, GRU) and kernel-based models (Rocket, Arsenal) were benchmarked.
Numeric Results:
- Digitization: Patched YOLOv11x achieved IoU=0.647 and Dice=0.782 on noisy PTB-XL. Overall pipeline achieved SNR=4.54 dB, Pearson r=0.806, and RMSE=0.043 mV on degraded sheets.
- Classification (PTB-XL MI vs. Normal): Best full-sequence accuracy: Arsenal (0.9583). Best segmented beat accuracy: Rocket (0.9551).
- Classification (ECG-Matrix Pre- vs. Post-procedural MI): Best accuracy: InceptionTime (0.9053).
- Classification (ECG-Matrix OMI vs. Non-OMI): Best accuracy: Rocket (0.8889).
Runtime/Training Cost:
- End-to-end pipeline inference: <30 s per ECG on CPU-only.
- YOLOv11x segmentation: ~0.3 s/image.
- Classification latency: Varies by model, ranging from ~0.26 ms (MLP) to several minutes (kernel methods).

The paper notes that for ECG-Matrix, paired waveform ground truth was not available, so digitization was evaluated indirectly through downstream diagnostic performance.
The pre-procedural vs. post-procedural comparison is described as an "exploratory proxy" that may reflect multiple factors beyond just the intervention.
The smaller sample size in ECG-Matrix may limit calibration and partially explain modest performance differences.
Failure modes in digitization can be driven by severe overlap between adjacent lead traces, missing calibration pulses, extreme baseline drift, or non-standard report layouts.

Method evidence: "We address this need with an end-to-end lightweight on-device digitization-to-diagnosis pipeline that converts a smartphone photo or scan of a paper ECG into a calibrated 12-lead signal and screens for Myocardial Infarction (MI) pathologies, with Shapley Additive explanations (SHAP) to support interpretability." (Abstract)
Result evidence: "On a CPU-only Intel Core i9-13900H laptop, the end-to-end pipeline runs in roughly 25-30 s per ECG." (Figure 1 caption)
Limitation evidence: "Consequently, for real-world scans we evaluate digitization indirectly through downstream diagnostic performance and interpretability rather than signal-level reconstruction error." (Section VI. DISCUSSION)

ECGLight: Compute-Light Framework For Paper ECG Digitization and Myocardial Infarction Screening

Abstract

Summary

Problem

Key Contributions

Method

Technical Details

Results

Weaknesses / Limitations

Additional Figures