Multi-label arrhythmia classification from Holter scatter plots using deep learning and clinically oriented thresholds

In a retrospective single-center study, researchers developed a deep learning framework to classify multiple arrhythmia labels from Holter scatter plot images. The study included 600 de-identified patients, each with one Holter record. Expert-reviewed rhythm conclusions were mapped to 15 rhythm labels for model evaluation purposes.

The researchers used ResNet18-based models, trained with predefined scatter plot crops. The primary model, trained with weighted binary cross-entropy, achieved a macro-F1 score of 0.738 (95% CI, 0.666–0.774), micro-F1 of 0.808 (0.777–0.838), macro-AUROC of 0.919 (0.895–0.941), and macro-AUPRC of 0.805 (0.758–0.858). This was achieved using a fixed threshold of 0.5 for classification.

However, applying clinically oriented thresholds increased mean specificity from 0.844 (0.822–0.865) to 0.885 (0.860–0.909). While specificity improved, the macro-F1 score decreased to 0.693 (0.628–0.726) due to the threshold adjustments. The study found strong discrimination for atrial fibrillation, atrial flutter, premature atrial contractions, and premature ventricular contractions.

The authors conclude that scatter plot-based deep learning may offer an efficient preliminary screening approach for Holter analysis. However, they emphasize the need for multicenter external validation before potential clinical implementation. The research received no specific funding and was conducted at The Second Affiliated Hospital of Anhui Medical University and the Institute of Intelligent Machines, Chinese Academy of Sciences.