BACKGROUND: Digitization and computerised analysis of ECG images allows researchers to derive new insights from historic records and overcomes issues arising from proprietary signal formats, but digitized signals suffer from a worse signal to noise ratio (SNR) than raw signals and can produce inferior results. Deep learning (DL) models can be particularly robust to poor SNRs and we propose that they may provide exceptionally robust analysis of image-derived ECG signals. We aim to benchmark image-based ECG analysis against raw signal analysis on a particularly challenging, ambulatory ECG task. METHODS: The 2017 Physionet AF Challenge data was downloaded and ECG records were plotted into human-readable images at 50% resolution. The ECG images were then digitized using established techniques. The image-derived signals were used to train a deep convolutional neural network for a four-class rhythm recognition task. Results from 5-fold cross validation on the public training set were directly compared with results from leading competition scorers, who used raw signals. RESULTS: A combined F1 score of 0.78 was obtained. This represents minimal performance loss compared with raw signal analysis, where six top competitors attained a mean combined F1 score of 0.83. CONCLUSION: Our findings show that DL-based analysis of ECG images is particularly robust. We advocate for a renewed interest in ECG image analysis using a DL approach.
|Journal||Journal of Electrocardiology|
|Publication status||Published (in print/issue) - 18 Oct 2019|
- Deep Learning
- Machine Learning
- heart attacks