ECG-FM is a foundation model for electrocardiogram (ECG) analysis. Committed to open-source practices, ECG-FM was developed in collaboration with the fairseq_signals framework, which implements a collection of deep learning methods for ECG analysis. This repository serves as a landing page and will host project-specific scripts as this work progresses.
- 2024-08-12: ECG-FM arxiv & GitHub released
ECG-FM adopts the wav2vec 2.0 architecture and was pretrained using the W2V+CMSC+RLM (WCR) method. It has 311,940,352 parameters and was trained using 4 NVIDIA A100 80GB GPUs over 16.5 days. For our transformer encoder, we selected hyperparameters consistent with a BERT-Large encoder. Further details are available in our paper.
We are committed to open-weight practices. Model checkpoints have been made publicly available for download on HuggingFace.
Specifically, there is:
mimic_iv_ecg_physionet_pretrained.pt
- Was pretrained on MIMIC-IV-ECG v1.0 and PhysioNet 2021 v1.0.3.
physionet_finetuned.pt
- Was finetuned from
mimic_iv_ecg_physionet_pretrained.pton PhysioNet 2021 v1.0.3.
Disclaimer: These models are different from those reported in our arXiv paper. These BERT-Base sized models were trained purely on public data sources due to privacy concerns surrounding UHN-ECG data and patient identification. Validation for the final models will be available upon full publication.
Clone fairseq_signals and refer to the requirements and installation section in the top-level README.
We implemented a flexible, end-to-end, multi-source data preprocessing pipeline. Please refer to it here.
See our inference tutorial notebook!
Training is performed through the fairseq_signals framework. To maximize reproducibility, we have provided configuration files.
Pretraining can be performed by downloading the mimic_iv_ecg_physionet_pretrained.yaml config (or modifying fairseq-signals/examples/w2v_cmsc/config/pretraining/w2v_cmsc_rlm.yaml as desired).
After modifying the relevant configuration file as desired, pretraining is performed using hydra's command line interface. This command highlights some popular config overrides:
FAIRSEQ_SIGNALS_ROOT="<TODO>"
MANIFEST_DIR="<TODO>/cmsc"
OUTPUT_DIR="<TODO>"
fairseq-hydra-train \
task.data=$MANIFEST_DIR \
dataset.valid_subset=valid \
dataset.batch_size=64 \
dataset.num_workers=10 \
dataset.disable_validation=false \
distributed_training.distributed_world_size=4 \
optimization.update_freq=[2] \
checkpoint.save_dir=$OUTPUT_DIR \
checkpoint.save_interval=10 \
checkpoint.keep_last_epochs=0 \
common.log_format=csv \
--config-dir $FAIRSEQ_SIGNALS_ROOT/examples/w2v_cmsc/config/pretraining \
--config-name w2v_cmsc_rlm
Classification finetuning uses the physionet_finetuned.yaml or fairseq-signals/examples/w2v_cmsc/config/finetuning/ecg_transformer/diagnosis.yaml configs. This command highlights some popular config overrides:
FAIRSEQ_SIGNALS_ROOT="<TODO>"
PRETRAINED_MODEL="<TODO>"
MANIFEST_DIR="<TODO>"
LABEL_DIR="<TODO>"
OUTPUT_DIR="<TODO>"
NUM_LABELS=$(($(wc -l < "$LABEL_DIR/label_def.csv") - 1))
POS_WEIGHT=$(cat $LABEL_DIR/pos_weight.txt)
fairseq-hydra-train \
task.data=$MANIFEST_DIR \
model.model_path=$PRETRAINED_MODEL \
model.num_labels=$NUM_LABELS \
optimization.lr=[1e-06] \
optimization.max_epoch=140 \
dataset.batch_size=256 \
dataset.num_workers=5 \
dataset.disable_validation=true \
distributed_training.distributed_world_size=1 \
distributed_training.find_unused_parameters=True \
checkpoint.save_dir=$OUTPUT_DIR \
checkpoint.save_interval=1 \
checkpoint.keep_last_epochs=0 \
common.log_format=csv \
+task.label_file=$LABEL_DIR/y.npy \
+criterion.pos_weight=$POS_WEIGHT \
--config-dir $FAIRSEQ_SIGNALS_ROOT/examples/w2v_cmsc/config/finetuning/ecg_transformer \
--config-name diagnosis
Notes:
- With CMSC pretraining, the batch size refers to pairs of adjacent segments. Therefore, the effective pretraining batch size is
64 pairs * 2 segments per pair * 4 GPUs * 2 gradient accumulations (update_freq) = 1024 segments. - ECG-FM has 311,940,352 parameters, whereas the base model has 90,883,072 parameters. We would not suggest pretraining a large model having only those public data sources (PhysioNet 2021 and MIMIC-IV-ECG) used in the paper.
Functionality for our comphensive free-text pattern matching and knowledge graph based label manipulation will be made available soon!
Inquiries may be directed to kaden.mckeen@mail.utoronto.ca.