QECGBERT tokenizes ECG waveforms for BERT-style pretraining
Original title: QECGBERT: A Scalable ECG Foundation Model Based on Self-Supervised Language Learning via VQ-VAE Waveform Tokenization
It is pretrained on more than 800,000 unlabeled MIMIC-IV ECGs and fine-tuned on three labeled datasets.
QECGBERT is a self-supervised ECG language model that tokenizes P, QRS, T, and isoelectric background waveforms rather than treating traces as generic sequences or images. A VQ-VAE tokenizes each 12-lead ECG and adds continuous waveform features so morphology is kept without longer sequences. BERT masked language modeling is pretrained on more than 800,000 unlabeled MIMIC-IV recordings, then fine-tuned on PTB-XL, CPSC2018, and Chapman-Shaoxing-Ningbo. The authors report comparable or better results than published benchmarks, aimed at cardiac diagnosis including rare diseases with little labeled data.
Unconfirmed: Comparable or superior benchmark performance is the authors' claim and is not independently confirmed here.
- Tokens cover P, QRS, T, and isoelectric background waveforms.
- Continuous features are added so quantization does not drop morphology or lengthen sequences.
- Pretraining uses the MIMIC-IV set of over 800,000 unlabeled ECG recordings.
- Evaluation fine-tuning uses PTB-XL, CPSC2018, and Chapman-Shaoxing-Ningbo.
- AI models
- QECGBERT
From the original article
1 min read at arxiv.orgAI foundation models play a pivotal role in advancing heart disease diagnosis from electrocardiogram (ECG) data, particularly for rare diseases where the lack of sufficient labeled data hinders supervised learning. However, most self-supervised learning methods treat ECGs as generic sequence or image data without exploiting underlying physiological structure.
By Aruna Mohan · arxiv.org
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.