Cpsc2018 Ecg Classification Eval

---
name: cpsc2018-ecg-classification-eval
description: Evaluates deep learning models on multi-lead ECG signal classification for arrhythmia detection under artificially balanced conditions. It probes the model's ability to extract discriminative temporal-spatial features from raw 12-lead cardiac signals and maintain robustness against various types of physiological noise. Use when the user wants to benchmark on CPSC2018, or asks about evaluating this task. Reports Accuracy.
metadata:
  skill_kind: dataset_eval
  source_arxiv: 2601.09103
  bibtex_key: shao2026enhancing
  confidence: medium
---

# cpsc2018-ecg-classification-eval

> Enhancing Imbalanced Electrocardiogram Classification: A Novel Approach Integrating Data Augmentation through Wavelet Transform and Interclass Fusion — Shao et al. (2026) (arXiv:2601.09103, 2026)

## What this evaluates

Evaluates deep learning models on multi-lead ECG signal classification for arrhythmia detection under artificially balanced conditions. It probes the model's ability to extract discriminative temporal-spatial features from raw 12-lead cardiac signals and maintain robustness against various types of physiological noise.

## Datasets

- **CPSC2018** — total 1917; splits: train (-1), val (-1), test (-1); repo https://github.com/Harmenlv/ECG_CPSC_2018

## Metrics

- `Accuracy` **(primary)** — range: [0, 1]
  - Standard classification accuracy: (Number of correct predictions) / (Total number of predictions). Reported as a fraction or percentage.
- `Precision` — range: [0, 1]
  - Ratio of true positive predictions to all positive predictions for each class, typically macro-averaged in multi-class settings.
- `Recall` — range: [0, 1]
  - Ratio of true positive predictions to all actual positives for each class, typically macro-averaged in multi-class settings.
- `AUC` — range: [0, 1]
  - Area Under the Receiver Operating Characteristic Curve. Computed using one-vs-rest strategy across all 9 classes and macro-averaged.

## Input / output format

**Input**: 12-lead ECG signal reshaped to a tensor of shape (12, 5000), representing 12 channels with 5000 time steps each.

**Output**: Single-class label from 9 categories: AF, I-AVB, LBBB, PAC, PVC, RBBB, STD, STE, Normal.

## Scoring recipe

```python
def compute_metrics(y_true, y_pred):
    accuracy = np.sum(y_true == y_pred) / len(y_true)
    precision = precision_score(y_true, y_pred, average='macro')
    recall = recall_score(y_true, y_pred, average='macro')
    auc = roc_auc_score(y_true, y_pred, multi_class='ovr', average='macro')
    return {'accuracy': accuracy, 'precision': precision, 'recall': recall, 'auc': auc}
```

## Common pitfalls

- The dataset is artificially balanced by subsampling majority classes to exactly 213 samples each, which masks the severe real-world class imbalance of the original CPSC2018 challenge.
- Noise robustness is evaluated by adding 20 specific dB levels (-27dB to 12dB), but the paper does not specify whether noise is injected during training, validation, or only testing, nor how metrics are aggregated across noise levels.
- Train/validation/test split sizes and ratios are never explicitly stated, making exact reproducibility of the evaluation split difficult.

## Evidence (verbatim from paper)

> In this experiment, a comprehensive set of evaluation metrics was employed to assess the performance of the models. Accuracy, loss, recall, precision, receiver operating characteristic (ROC), and area under the curve (AUC) were selected as the evaluation metrics for this study.

## Citation

```bibtex
@misc{shao2026enhancing,
  title={Enhancing Imbalanced Electrocardiogram Classification: A Novel Approach Integrating Data Augmentation through Wavelet Transform and Interclass Fusion},
  author={Shao et al. (2026)},
  year={2026},
  note={arXiv:2601.09103}
}
```

- arXiv: 2601.09103