Image based deep learning in 12-lead ECG diagnosis
Problem
In many clinical settings, especially remote and regional ones, ECGs exist only as printed paper or digital images, while most machine learning diagnostic tools require raw signal data.
Why it matters
The ECG is an integral tool for diagnosing cardiovascular disease, so image-only formats limit automated diagnosis in many areas.
What the AI did
VGG-architecture deep learning models were trained on 12-lead ECG image datasets and evaluated on holdout test data and datasets not seen in training, with Grad-CAM heatmaps depicting diagnoses.
Results
Stated qualitatively: excellent AUROC, AUPRC, sensitivity and specificity on holdout data, comparable to the best signal and image-based models; no numeric values in abstract.
Impact
Shows image-based deep learning can diagnose 12-lead ECGs at performance comparable to the best existing models.
Record
| Venue | Frontiers in Artificial Intelligence |
|---|---|
| Published | 2023 |
| Work type | article |
| Authors | Raymond Ao, George He |
| Citations | 41 (OpenAlex, at import time) |
| References | 28 |
| Model or method | VGG architecture (with Grad-CAM) |
| Organization | not stated in source |
| Primary source | https://doi.org/10.3389/frai.2022.1087370 |
| Open access | gold |
| Keywords | deep learning, ECG images, VGG, Grad-CAM, cardiovascular diagnosis |
Importance rating
significant — thousands to millions affected
Abstract states ECG is an integral tool for diagnosing cardiovascular disease and that image-only access limits ML diagnosis in many clinical settings, especially remote/regional ones.