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.

SSRL-ECG uses two publicly available PhysioNet datasets. PTB-XL is the primary training and evaluation dataset — a large-scale 12-lead ECG library with standardized diagnostic labels and a pre-defined 10-fold stratified split. MIT-BIH Arrhythmia Database serves as an independent external benchmark to test how well representations learned on PTB-XL transfer to a different recording protocol and patient population.
The package loader expects a specific folder hierarchy under your data/ directory. Placing files in a different layout will cause FileNotFoundError at runtime. See the required folder structure below before downloading.

​
PTB-XL — Primary Dataset

PTB-XL is a 12-lead clinical ECG database collected at the Physikalisch-Technische Bundesanstalt in Berlin. It is the largest openly available annotated 12-lead ECG dataset and forms the backbone of all SSRL-ECG pretraining, fine-tuning, and evaluation experiments.

​
Recording Specifications

PropertyValue
Total ECGs21,799
Unique patients18,869
Leads12 (standard clinical layout)
Sampling rate500 Hz (high-res) / 100 Hz (low-res)
Recording duration10 seconds per ECG
Signal length at 500 Hz5,000 samples per lead

​
Diagnostic Superclasses

PTB-XL labels are organized into five mutually non-exclusive cardiovascular disease superclasses. Each ECG can carry one or more labels:
ClassFull nameSample count
NORMNormal ECG9,514
MIMyocardial Infarction5,469
STTCST/T-wave Changes5,235
CDConduction Disturbance4,898
HYPLeft Ventricular Hypertrophy2,649
The class imbalance ratio is 3.32× (NORM vs. HYP). The supervised baseline addresses this with focal loss and oversampling. The SSL pretraining phase does not use labels, so class imbalance does not directly affect the contrastive or momentum objective — but it does affect the downstream linear probe.

​
10-Fold Split

PTB-XL ships with a pre-assigned strat_fold column (1–10) for reproducible stratified splitting. SSRL-ECG maps these folds to fixed train / validation / test partitions:
SplitFoldsSamples
Train1–817,489
Validation92,154
Test102,194
The label-efficient fine-tuning experiment draws 1,747 labeled samples — exactly 10% of the training split — while the SSL pretraining uses all 17,489 training recordings without labels.

​
MIT-BIH Arrhythmia Database — External Validation

The MIT-BIH Arrhythmia Database provides 48 half-hour two-lead ambulatory ECG recordings from 47 subjects, sampled at 360 Hz with beat-level arrhythmia annotations. In SSRL-ECG it is used exclusively as an external robustness check: representations pretrained on PTB-XL are evaluated on MIT-BIH data to assess domain generalization without any fine-tuning on MIT-BIH itself.
PropertyValue
Records48
Annotation typeBeat-level arrhythmia labels
Leads2 (modified limb lead II + V-lead)
Sampling rate360 Hz
Record duration~30 minutes each
Use in SSRL-ECGExternal transfer / robustness validation

​
Required Folder Structure

Both datasets must be organized under a single data/ directory at your project root. The PTB-XL loader reads ptbxl_database.csv for metadata and scp_statements.csv for diagnostic code mappings; the signal files are located via the filename_lr (100 Hz) and filename_hr (500 Hz) columns in the CSV. 
data/
├── PTB-XL/
│   ├── ptbxl_database.csv        # metadata + fold assignments
│   ├── scp_statements.csv        # diagnostic code → superclass mapping
│   ├── records100/               # 100 Hz downsampled recordings
│   │   ├── 00000/
│   │   │   ├── *.hea             # WFDB header files
│   │   │   └── *.dat             # raw signal data
│   │   └── ...
│   └── records500/               # 500 Hz full-resolution recordings
│       ├── 00000/
│       │   ├── *.hea
│       │   └── *.dat
│       └── ...
└── MIT-BIH/
    └── files/
        └── mitdb/
            └── 1.0.0/
                ├── *.hea         # WFDB header files
                ├── *.dat         # raw signal data
                └── *.atr         # beat annotation files
PTB-XL ships signal files at two resolutions. SSRL-ECG defaults to 500 Hz (records500/) for all pretraining and fine-tuning experiments. The records100/ directory (100 Hz) is available for faster prototyping or resource-constrained environments — signal length becomes 1,000 samples at 100 Hz instead of 5,000 at 500 Hz. Pass --signal-length 1000 to training scripts when using the lower-resolution records.

​
Downloading the Datasets

PTB-XL on PhysioNet

Download PTB-XL v1.0.3 from PhysioNet. Requires a free PhysioNet account. The full dataset is ~2.1 GB compressed.

MIT-BIH on PhysioNet

Download MIT-BIH Arrhythmia Database v1.0.0 from PhysioNet. Approximately 106 MB compressed.
After downloading, unzip both archives and arrange the files to match the folder structure shown above. PTB-XL’s archive already contains the records100/ and records500/ subdirectories — move the top-level folder to data/PTB-XL/. For MIT-BIH, create the data/MIT-BIH/files/mitdb/1.0.0/ path and place all .hea, .dat, and .atr files directly inside it.

​
Verifying Your Dataset Setup

Run the built-in analysis script to confirm both datasets are correctly placed and print class distribution statistics: 
python -m ssrl_ecg.analyze_datasets --ptbxl-root data/PTB-XL
Expected output: 
PTB-XL rows: 21799
PTB-XL unique patients: 18869
PTB-XL fold sizes: {'train(1-8)': 17489, 'val(9)': 2154, 'test(10)': 2194}
PTB-XL superclass counts:
  NORM: 9514
  MI: 5469
  STTC: 5235
  HYP: 2649
  CD: 4898
PTB-XL availability: {'lr_dat': '21799/21799', 'hr_dat': '21799/21799'}
MIT-BIH availability: {'hea': 48, 'dat': 48, 'atr': 48}
The availability lines confirm that all signal .dat files referenced in ptbxl_database.csv are present on disk, and that all 48 MIT-BIH records have their header, data, and annotation files. Any count below the expected value indicates missing files that need to be re-downloaded.
If lr_dat or hr_dat availability is below 21799/21799, the dataset loader will silently skip missing records during training, which can skew class distributions in the training batches. Always verify 100% file availability before starting a pretraining run.

Build docs developers (and LLMs) love

Get started for freeTalk to us