Pacaf Ecg Classification Eval

pacaf-ecg-classification-eval

Accurate Detection of Paroxysmal Atrial Fibrillation with Certified-GAN and Neural Architecture Search — Asadi et al. (2023) (arXiv:2301.10173, 2023)

What this evaluates

Evaluates a model's ability to classify 4-second two-channel ECG segments as either normal (healthy) or paroxysmal atrial fibrillation (PAxF). It probes the model's diagnostic accuracy and sensitivity in detecting cardiac arrhythmia from raw physiological signals.

Datasets

• PhysioNet PxAF prediction challenge database — total 6043; splits: train (4231), val (906), test (906)

Metrics

• Accuracy (primary) — range: [0, 1]
  • Accuracy = (TP + TN) / (TP + TN + FP + FN), where TP, TN, FP, and FN denote True Positives, True Negatives, False Positives, and False Negatives, respectively.
• Specificity — range: [0, 1]
  • Specificity = TN / (TN + FP), measuring the proportion of actual negatives correctly identified.
• Sensitivity — range: [0, 1]
  • Sensitivity = TP / (TP + FN), measuring the proportion of actual positives correctly identified.

Input / output format

Input: 4-second two-channel ECG signal segments sampled at 128 Hz (16-bit resolution).

Output: Binary classification label: 'normal' (healthy) or 'PAxF' (paroxysmal atrial fibrillation).

Scoring recipe

def compute_metrics(tp, tn, fp, fn):
    accuracy = (tp + tn) / (tp + tn + fp + fn)
    specificity = tn / (tn + fp)
    sensitivity = tp / (tp + fn)
    return accuracy, specificity, sensitivity

Common pitfalls

• The training set is augmented with 539 expert-certified synthetic GAN segments, which are not part of the original PhysioNet split and must be explicitly added to reproduce the reported results.
• The dataset uses 4-second segmented windows derived from 30-minute recordings, not the raw recordings themselves, which significantly alters the effective sample size and class distribution.
• Despite creating a 906-segment validation split, the authors explicitly state they did not use a separate validation set to evaluate training performance due to the small database size.

Evidence (verbatim from paper)

The formulas for quantifying measurements are listed below:

$$ A c c u r a c y = rac {T P + T N}{T P + T N + F P + F N} ag {4} $$

$$ S p e c i f i c i t y = rac {T N}{T N + F P} $$

$$ S e n s i t i v i t y = rac {T P}{T P + F N} $$ where $TP, TN, FP$ , and $FN$ denote True Positives, True False Positive, and False Negative, respectively.

Citation

@misc{asadi2023accurate,
  title={Accurate Detection of Paroxysmal Atrial Fibrillation with Certified-GAN and Neural Architecture Search},
  author={Asadi et al. (2023)},
  year={2023},
  note={arXiv:2301.10173}
}

• arXiv: 2301.10173