Mimic Iv Ecg Classification Eval

---
name: mimic-iv-ecg-classification-eval
description: This evaluation probes the ability of transformer-based models to classify cardiac rhythms using time-series features derived from dynamical systems theory (Koopman operator) and signal processing (wavelets). It specifically tests binary and four-class rhythm classification on clinical ECG waveforms, comparing feature-based approaches against raw RNN baselines. Use when the user wants to benchmark on MIMIC-IV-ECG, or asks about evaluating this task. Reports F1 score.
metadata:
  skill_kind: dataset_eval
  source_arxiv: 2603.08339
  bibtex_key: [REDACTED]
  confidence: high
---

# mimic-iv-ecg-classification-eval

> Electrocardiogram Classification with Transformers Using Koopman and Wavelet Features — Ghosh et al. (2026) (arXiv:2603.08339, 2026)

## What this evaluates

This evaluation probes the ability of transformer-based models to classify cardiac rhythms using time-series features derived from dynamical systems theory (Koopman operator) and signal processing (wavelets). It specifically tests binary and four-class rhythm classification on clinical ECG waveforms, comparing feature-based approaches against raw RNN baselines.

## Datasets

- **MIMIC-IV-ECG** — total ?; splits: train (-1), val (-1), test (-1)

## Metrics

- `F1 score` **(primary)** — range: [0, 1]
  - Harmonic mean of precision and recall, computed on the held-out test set. Reported as mean ± standard deviation over 5 independent runs. Calculated per task (binary and 4-class).

## Input / output format

**Input**: Fixed-length, z-score normalized ECG windows (50% overlap segmentation) with features extracted via wavelet decomposition or Koopman mode decomposition before being fed to the transformer.

**Output**: Predicted diagnostic label (binary: Normal/Non-normal or 4-class: Normal/AFib/Ventricular Arrhythmias/Conduction Block).

## Scoring recipe

```python
def compute_f1(predictions, gold):
    # predictions, gold: lists of class labels
    # Compute macro-averaged F1 score across all classes
    from sklearn.metrics import f1_score
    return f1_score(gold, predictions, average='macro')
```

## Common pitfalls

- Combining Koopman and wavelet features (hybrid fusion) unexpectedly degrades performance rather than improving it.
- Koopman-based classification requires careful tuning of the radial basis function dictionary; naive application underperforms wavelets in binary settings.
- Raw RNN baselines on raw ECG signals are computationally prohibitive at scale, making feature-based approaches necessary for practical deployment.

## Evidence (verbatim from paper)

> The performance was evaluated using the F1 score on the held-out test set. We observed that the raw RNN baseline is computationally prohibitive at larger scale, whereas Koopman- and wavelet-based feature methods are much more efficient.

## Citation

```bibtex
@misc{ghosh2026ecgkoopmanwavelet,
  title={Electrocardiogram Classification with Transformers Using Koopman and Wavelet Features},
  author={Ghosh et al. (2026)},
  year={2026},
  note={arXiv:2603.08339}
}
```

- arXiv: 2603.08339