It is our pleasure to annonce the publication of our article “RawECGNet: Deep Learning Generalization for Atrial Fibrillation Detection from the Raw ECG” published in IEEE Journal of Biomedical and Health Informatics (JBHI). URL: https://ieeexplore.ieee.org/document/10538381
Considerable research has been devoted to developing methods for atrial fibrillation (AF) detection, with the aim of increasing the detection accuracy and reducing the workload of clinical staff. AF detection has evolved from methods using rhythm features and morphological features, followed by thresholding or machine learning-based classification, to arrive at today's large variety of deep learning-based detectors. Deep learning-based AF detectors designed to process single-lead ECGs can be broadly categorized according to whether the beat-to-beat interval or the raw ECG serve as input to the network. The latter category is theoretically richer in information although this has not been rigorously and quantitatively proven.
To achieve high performance, it is crucial that deep learning models be robust, meaning that the models should generalize across important distribution shifts to ensure its efficacy when deployed in a variety of health settings and population samples. Indeed, although high performance if often reported for deep learning models using raw physiological signals, the models tend to generalize poorly or moderately well when evaluated on other data sets. Previous research has not evaluated the generalization performance of raw ECG-based AF detectors across important sources of distribution shift including the lead position, geography, sex and age.
Addressing these research gaps, our research makes the following contributions:
- RawECGNet, a new deep learning model for AF detection based on the raw, single-lead ECG; In particular, the architecture and strategy that we use to train RawECGNet and create a generalizable representation of the raw ECG data towards AF detection.
- Rigorous benchmarking of RawECGNet against a state-of-the-art, rhythm information-based model ArNet2;
- Evaluation of generalization performance on two manually annotated independent data sets from different geographical locations as well as across lead position, age, and sex; the data sets, total over 5000 hours of continuous physiological data.
- A thorough quantitative error analysis pinpointing the main sources of false positives and false negatives.
In conclusion, we developed a deep learning model, named RawECGNet, for the detection of episodes of atrial fibrillation and atrial flutter from the raw, single-lead ECG. We have demonstrated the generalization of RawECGNet on two external data sets that account for distribution shifts in geography, ethnicity, and lead position.
Figure: Single channel ECGs are used to train the deep learning model RawECGNet for the task of atrial fibrillation events detection in long term recordings. Training utilizes data from various lead positions for broad representation. RawECGNet generalization performance is assessed on datasets from diverse geographical areas and across lead positions, ages, and sex. Article URL: https://ieeexplore.ieee.org/document/10538381
Comentários