SafeECGMatch: Calibration-Aware Joint Frequency and Time Space Semi-Supervised Learning for Open-Set ECG Classification

Hongkyu Koh    Ikbeom Jang ^†
† Corresponding author

Official repository for the paper

SafeECGMatch: Calibration-Aware Joint Frequency and Time Space Semi-Supervised Learning for Open-Set ECG Classification

Abstract

Electrocardiogram (ECG) classification models often suffer from severe label scarcity, making semi-supervised learning (SSL) an at- tractive strategy for reducing annotation costs. In clinical settings, however, unlabeled pools frequently contain out-of-distribution (OOD) anomalies or diagnostic groups absent from the labeled set. Standard SSL forces incorrect pseudo-labels onto these unseen classes, producing overconfident predictions. To address this, we propose SafeECGMatch, a calibration-aware safe SSL framework for single-label ECG classification under label distribution mis- match. Methodologically, SafeECGMatch employs a dual-branch architecture extracting time-frequency latent representations via ECG-specific augmentations. Crucially, it dynamically aligns confi- dence with empirical accuracy through adaptive label smoothing and temperature scaling, calibrating both the multiclass classifier and the OOD detector across temporal and spectral domains. This joint optimization allows trustworthy OOD rejection and reliable pseudo-labeling. Evaluated on the PTB-XL and PhysioNet/CinC Challenge benchmarks, SafeECGMatch achieves state-of-the-art accuracy and calibration, advancing reliable knowledge discovery in physiological time-series.

What Is Here

• scripts/run_paper_benchmarks.py: main entrypoint for the benchmark suites
• scripts/preprocess_cinc2021.py: CINC2021 preprocessing into the release-ready format
• scripts/run_safeecgmatch_sensitivity.py: SafeECGMatch branch-weight sensitivity runner
• configs.py, datasets/, main/, models/, tasks/, utils/: ECG runtime code used by the release
• resources/cinc_labels/*.txt: superclass label mappings used for CINC2021 preprocessing

The supported release surface is limited to ECG datasets (PTB-XL, Chapman, Georgia, Ningbo, CINC2021) with the resnet1d backbone.

Dataset Download and Storage

Both PTB-XL and CINC2021 should be downloaded from their official PhysioNet releases.

• PTB-XL: download the official PTB-XL release and extract it locally.
• CINC2021: download the PhysioNet Challenge 2021 training data and keep the raw training directory locally.

Recommended local layout:

/path/to/ecg-data/
  ptb-xl/
    1.0.3/
  challenge-2021/
    1.0.3/
      training/
  cinc2021_single_label_processed/

Path rules used by this release:

• --ptbxl-root should point to the extracted PTB-XL root, for example /path/to/ecg-data/ptb-xl/1.0.3.
• --cinc2021-root should point to the processed directory created by scripts/preprocess_cinc2021.py, not the raw Challenge directory.
• The exact storage location is up to the user. These paths do not need to match our server.

The smoke tests used during release preparation succeeded because these datasets were already available on this server. On another machine, the same commands will work as long as the user passes their own local dataset paths.

Quick Start

1. Install

pip install -r requirements.txt

Extra preprocessing dependencies are already included in requirements.txt.

2. Choose Your Entry Point

• Use scripts/run_paper_benchmarks.py for the main paper benchmarks.
• Use scripts/run_safeecgmatch_sensitivity.py for the SafeECGMatch sensitivity study.
• Use scripts/preprocess_cinc2021.py only when you need to build the processed CINC2021 release dataset.

3. Run PTB-XL Benchmarks

python scripts/run_paper_benchmarks.py \
  --benchmarks ptbxl_30_ood ptbxl_60_ood \
  --ptbxl-root /path/to/ptb-xl/1.0.3

4. Run CINC2021 Benchmarks

Preprocess once:

python scripts/preprocess_cinc2021.py \
  --source-root /path/to/cinc2021/raw \
  --output-root /path/to/cinc2021_single_label_processed

Then run the benchmark suites:

python scripts/run_paper_benchmarks.py \
  --benchmarks cinc2021_30_ood cinc2021_60_ood \
  --cinc2021-root /path/to/cinc2021_single_label_processed

5. Run SafeECGMatch Sensitivity

python scripts/run_safeecgmatch_sensitivity.py \
  --variants freqheavy timeheavy \
  --ptbxl-root /path/to/ptb-xl/1.0.3

Datasets

PTB-XL

Pass the official raw PTB-XL root directly through --ptbxl-root.

Example root:

/path/to/ecg-data/ptb-xl/1.0.3

CINC2021

The release expects a processed CINC2021 directory produced by scripts/preprocess_cinc2021.py.

Workflow:

1. Download the raw Challenge 2021 training data.
2. Store it anywhere locally, for example /path/to/ecg-data/challenge-2021/1.0.3/training.
3. Run scripts/preprocess_cinc2021.py once.
4. Pass the resulting processed directory through --cinc2021-root.

Expected output structure:

• metadata_single_label.csv
• data/{id}.npy
• preprocess_summary.json

The preprocessing follows the ECGMatch superclass mapping and keeps only single-group samples. The special case 426783006 is treated as Normal only when it is the sole diagnosis code.

Benchmark Suites

• ptbxl_30_ood: PTB-XL, 500 Hz, 30% OOD, full method suite
• cinc2021_30_ood: CINC2021, 500 Hz, 30% OOD, TS_TFC and CompleMatch
• ptbxl_60_ood: PTB-XL, 500 Hz, 60% OOD, TS_TFC and CompleMatch
• cinc2021_60_ood: CINC2021, 500 Hz, 60% OOD, TS_TFC and CompleMatch, with cinc-id-classes = [Rhythm, CD, Other] and cinc-ood-classes = [Normal, ST]

Legacy numeric aliases 05, 06, 07, and 08 are still accepted, but the descriptive names above are the supported release interface.

Common options:

• --seeds 1 2 3: override the default seeds
• --gpus 0: choose GPU ids passed through to the training scripts
• --dry-run: print commands without executing them
• --collect-only: skip execution and only aggregate metrics from completed runs

Results are written under checkpoints/ and aggregated summaries are written under results/.

Sensitivity Variants

• freqheavy: lambda-time-branch = 0.5, lambda-freq-branch = 1.5
• timeheavy: lambda-time-branch = 1.5, lambda-freq-branch = 0.5

Both variants also use lambda-ova-cali = 0.1 and lambda-ova = 0.1 on top of the ptbxl_60_ood benchmark settings.

Notes

• The release scripts do not require NAS-specific paths. Dataset roots are passed explicitly with CLI flags.
• The GitHub-facing entrypoints are the scripts under scripts/; most users should not need to call files under main/ directly.