Ecgbench Eval

---
name: ecgbench-eval
description: Evaluates multimodal LLMs on interpreting electrocardiogram (ECG) images across classification, clinical report generation, and open-ended QA tasks. Probes robustness to real-world image artifacts, out-of-domain generalization, and clinical reasoning capabilities. Use when the user wants to benchmark on ECGBench, or asks about evaluating this task. Reports Accuracy.
metadata:
  skill_kind: dataset_eval
  source_arxiv: 2410.19008
  bibtex_key: liu2024teachmultimodal
  confidence: high
---

# ecgbench-eval

> Teach Multimodal LLMs to Comprehend Electrocardiographic Images — Liu et al. (2024) (arXiv:2410.19008, 2024)

## What this evaluates

Evaluates multimodal LLMs on interpreting electrocardiogram (ECG) images across classification, clinical report generation, and open-ended QA tasks. Probes robustness to real-world image artifacts, out-of-domain generalization, and clinical reasoning capabilities.

## Datasets

- **ECGBench** — total ?; splits: in-domain (-1), out-of-domain (-1)

## Metrics

- `AUC` — range: [0, 1]
  - Area Under the Receiver Operating Characteristic Curve, measuring binary/multiclass classification performance across all thresholds.
- `F1` — range: [0, 1]
  - Harmonic mean of precision and recall, computed per class and averaged (typically weighted) for multi-class tasks.
- `Accuracy` **(primary)** — range: [0, 1]
  - Proportion of correctly predicted labels out of total instances.
- `Report Score` — range: other
  - Score for clinical report generation quality, evaluated via LLM or human judgment against ground-truth reports.
- `Arena Score` — range: other
  - Score for multi-turn open-ended clinical QA, derived from pairwise comparisons or LLM judging in a benchmark format.

## Input / output format

**Input**: ECG image paired with a text prompt or question in a chatbot-style multi-turn dialogue format, using the <image> token to inject visual features.

**Output**: Text response: class labels, structured clinical reports, or conversational answers.

## Scoring recipe

```python
def compute_ecgbench_metrics(preds, golds):
    # Classification tasks (PTB-XL Super, CODE-15%, CPSC, CSN, G12, MMMU ECG)
    acc = sum(p == g for p, g in zip(preds, golds)) / len(golds)
    auc = roc_auc_score(golds, preds) # expects probabilities or labels
    f1 = f1_score(golds, preds, average='weighted')
    
    # Report generation (PTB-XL Report)
    report_score = llm_judge_or_human_eval(preds, golds)
    
    # Multi-turn QA (ECG Arena)
    arena_score = pairwise_elo_or_llm_judge(preds, golds)
    
    return {'AUC': auc, 'F1': f1, 'Accuracy': acc, 'Report Score': report_score, 'Arena Score': arena_score}
```

## Common pitfalls

- Domain-specific baselines use different evaluation setups and are explicitly noted as reference-only.
- Models often over-rely on OCR of printed text/numbers in ECG images rather than analyzing visual waveform patterns.
- Report generation and multi-turn QA are significantly harder and require step-wise reasoning, which current models struggle with.

## Evidence (verbatim from paper)

> Specifically, PULSE surpasses the best proprietary model (GPT-4o) with a 27% improvement in AUC, an 11-point gain in report score, and a 39% increase in accuracy on the PTB-XL Super, PTB-XL Report, and ECG-QA tasks, respectively.

## Citation

```bibtex
@misc{liu2024teachmultimodal,
  title={Teach Multimodal LLMs to Comprehend Electrocardiographic Images},
  author={Liu et al. (2024)},
  year={2024},
  note={arXiv:2410.19008}
}
```

- arXiv: 2410.19008