Cl Drive Cognitive Load Eval

---
name: cl-drive-cognitive-load-eval
description: Evaluates a model's ability to classify cognitive load levels from raw, multimodal physiological signals (EEG, ECG, EDA) collected during driving scenarios. It probes the model's capacity to learn temporal and cross-modal patterns without hand-crafted features. Use when the user wants to benchmark on CL-Drive, or asks about evaluating this task. Reports accuracy.
metadata:
  skill_kind: dataset_eval
  source_arxiv: 2507.14163
  bibtex_key: qiu2025uniphy net
  confidence: high
---

# cl-drive-cognitive-load-eval

> UniPhyNet: A Unified Network For Multimodal Physiological Raw Signal Classification — Qiu et al. (2025) (arXiv:2507.14163, 2025)

## What this evaluates

Evaluates a model's ability to classify cognitive load levels from raw, multimodal physiological signals (EEG, ECG, EDA) collected during driving scenarios. It probes the model's capacity to learn temporal and cross-modal patterns without hand-crafted features.

## Datasets

- **CL-Drive** — total ?; splits: train (-1), val (-1), test (-1)

## Metrics

- `accuracy` **(primary)** — range: [0, 1]
  - Standard classification accuracy: the proportion of correctly predicted cognitive load labels (binary or ternary) out of the total number of 10-second windows in the validation/test fold.

## Input / output format

**Input**: 10-second windows of raw, preprocessed physiological signals: 4-channel EEG (256 Hz), 3-channel ECG (512 Hz), and 3-channel EDA (128 Hz), filtered and normalized per subject/session.

**Output**: Categorical label: 'low'/'high' for binary classification, or 'low'/'medium'/'high' for ternary classification.

## Scoring recipe

```python
def calculate_accuracy(predictions, gold_labels):
    correct = sum(1 for p, g in zip(predictions, gold_labels) if p == g)
    return correct / len(gold_labels)
```

## Common pitfalls

- Using random train/val splits instead of the specified Leave-One-Subject-Out (LOSO) 10-fold cross-validation, which inflates performance due to subject-specific signal correlations.
- Incorrectly mapping the original 1-9 cognitive load ratings to binary or ternary classes, as the paper uses specific thresholds (1-4 vs 5-9 for binary; 1-3, 4-6, 7-9 for ternary).
- Skipping the specified band-pass filtering and normalization steps, which significantly alter the raw signal distributions and affect model convergence.

## Evidence (verbatim from paper)

> We train all models using 10-fold, and leave one subject out (LOSO), cross-validation settings. We explore both binary and ternary classification of cognitive load. ... For binary classification, we group the cognitive load ratings from 1 to 4 as 'low' cognitive load and 5 to 9 as 'high' cognitive load. For ternary classification, we divide cognitive load ratings into 3 groups, 1 to 3, 4 to 6, and 7 to 9, which correspond to classes of cognitive load 'low', 'medium' and 'high', respectively.

## Citation

```bibtex
@misc{qiu2025uniphy net,
  title={UniPhyNet: A Unified Network For Multimodal Physiological Raw Signal Classification},
  author={Qiu et al. (2025)},
  year={2025},
  note={arXiv:2507.14163}
}
```

- arXiv: 2507.14163