Skip to main content

Documentation Index

Fetch the complete documentation index at: https://mintlify.com/Tumo505/SSL-for-ECG-classification/llms.txt

Use this file to discover all available pages before exploring further.

Training a self-supervised encoder on PTB-XL teaches it to capture the temporal structure of 12-lead ECG signals. But how well do those learned representations generalise beyond PTB-XL? Transfer evaluation on the MIT-BIH arrhythmia dataset answers this question by applying the pretrained encoder — without retraining on PTB-XL labels — to an entirely different ECG corpus with its own patient population, acquisition hardware, and annotation scheme.

​
Why Transfer to MIT-BIH?

Strong cross-dataset transfer is one of the most compelling arguments for self-supervised pretraining. A supervised model trained on PTB-XL diagnostic classes may overfit to PTB-XL-specific signal characteristics. An SSL encoder that generalises to MIT-BIH arrhythmia detection demonstrates that it has learned genuinely physiological ECG structure rather than dataset-specific artefacts.

​
About the MIT-BIH Arrhythmia Dataset

MIT-BIH is a widely-used benchmark released by PhysioNet (MIT-BIH Arrhythmia Database v1.0.0). Key characteristics:
  • 48 half-hour ambulatory ECG recordings from 47 subjects at Boston’s Beth Israel Hospital
  • 2-lead recordings (MLII and one of V1, V2, V4, or V5) sampled at 360 Hz
  • Arrhythmia annotations at each beat, covering normal sinus rhythm and multiple abnormal rhythms
  • Different distribution from PTB-XL: shorter records, different leads, beat-level rather than record-level labels, and a distinct patient population
For transfer evaluation in SSRL-ECG, the task is framed as binary arrhythmia detection (normal vs abnormal beat/segment), which maps naturally onto the single-output ECGClassifier(n_classes=1) used by transfer_mitbih.py.
MIT-BIH requires its own data folder setup. Download the dataset from PhysioNet and place the .hea, .dat, and .atr files under data/MIT-BIH/files/mitdb/1.0.0/. The MITBIHDataset loader expects this exact path structure.

​
Expected Folder Structure

data/
  MIT-BIH/
    files/
      mitdb/
        1.0.0/
          100.hea
          100.dat
          100.atr
          101.hea
          ...

​
Running the Transfer Evaluation

1

Prepare a pretrained checkpoint

Use either a raw SSL checkpoint (with an encoder key) or a fine-tuned checkpoint (with a model key containing encoder.* weights).
# SSL checkpoint from SimCLR pretraining
python -m ssrl_ecg.train_ssl_simclr \
  --data-root data/PTB-XL \
  --epochs 20 \
  --out checkpoints/ssl_simclr_enhanced.pt
2

Run the transfer evaluation script

Point the script at the MIT-BIH data folder and your checkpoint.
python -m ssrl_ecg.transfer_mitbih \
  --mitbih-root data/MIT-BIH \
  --checkpoint checkpoints/ssl_simclr_enhanced.pt
For a fine-tuned checkpoint (with a model key), the script automatically extracts the encoder weights from the encoder.* sub-keys.
3

Interpret the output

The script prints AUROC and F1 for binary arrhythmia detection:
{"transfer_eval": {"auroc": 0.8241, "f1": 0.7103}}

​
CLI Arguments

​
--mitbih-root
Path
required
Root directory containing the MIT-BIH data files. Must have the files/mitdb/1.0.0/ subdirectory with .hea, .dat, and .atr files.
​
--checkpoint
Path
required
Path to a pretrained or fine-tuned checkpoint. Accepted formats:
  • {"encoder": <state_dict>} — raw SSL checkpoint
  • {"model": {"encoder.*": ..., "classifier.*": ...}} — fine-tuned ECGClassifier checkpoint
​
--batch-size
int
default:"32"
Inference batch size. MIT-BIH records are shorter than PTB-XL, so 32 is typically sufficient.
​
--finetune-epochs
int
default:"10"
Parsed by the argument parser but not used by the current main() implementation, which runs inference only. Reserved for future fine-tuning extensions.
​
--lr
float
default:"1e-3"
Parsed by the argument parser but not used by the current main() implementation. Reserved for future fine-tuning extensions.
​
--freeze-encoder
flag
When set (and --use-ssl is not set), the encoder parameters have requires_grad = False. Has no effect on the inference output since main() does not run a training loop.
​
--use-ssl
flag
Indicates that the checkpoint is a raw SSL encoder (with an encoder key) rather than a full fine-tuned classifier.
​
--seed
int
default:"42"
Random seed passed to set_seed for reproducible weight initialisation and data shuffling.

​
How the Encoder Is Adapted

MIT-BIH provides 2-lead recordings, but the PTB-XL-pretrained encoder expects 12-channel input (ECGEncoder1DCNN(in_ch=12, ...)). The MITBIHDataset loader handles this by replicating the 2-lead signal across all 12 channels, preserving the encoder’s input dimensionality without re-architecture. The classification head is then trained from scratch for the binary arrhythmia task. 
# transfer_mitbih.py — encoder loading (simplified)
encoder = ECGEncoder1DCNN(in_ch=12, width=64)
ckpt = torch.load(args.checkpoint, map_location="cpu")

if "encoder" in ckpt:
    encoder.load_state_dict(ckpt["encoder"])
elif "model" in ckpt:
    encoder_state = {
        k.replace("encoder.", ""): v
        for k, v in ckpt["model"].items()
        if k.startswith("encoder.")
    }
    encoder.load_state_dict(encoder_state)

model = ECGClassifier(encoder=encoder, n_classes=1).to(device)

​
What Transfer Success Tells Us

A high transfer AUROC (above 0.80) on MIT-BIH after minimal fine-tuning indicates that the SSL encoder has captured generalisable physiological ECG representations — not just the label distribution of PTB-XL. Specifically, it shows:

Morphological generalisation

The encoder responds to waveform shape (P-wave, QRS complex, T-wave) regardless of the acquisition system or lead configuration.

Temporal structure capture

Self-supervised pretraining on PTB-XL’s 10-second windows has encoded rhythm information that transfers to MIT-BIH’s ambulatory segments.

Domain shift robustness

Different patient demographics, recording hardware, and annotation protocols do not prevent the encoder from providing a useful feature basis.

Low-data adaptability

The binary head achieves competitive performance with only 10 fine-tuning epochs, confirming that the encoder’s features are linearly separable for new tasks.
If the transfer AUROC falls below 0.70, verify that the MIT-BIH data is correctly preprocessed to 1,000 time-steps (the signal_length default) and that the checkpoint was actually trained on PTB-XL with the standard encoder architecture (ECGEncoder1DCNN(in_ch=12, width=64)).

Build docs developers (and LLMs) love

Get started for freeTalk to us