Deep learning-based classification of atrial fibrillation severity: computational approach to support personalized clinical management
Rion ISHIKUBO *1, Takahashi YUYA2, 3, Yamaguchi TAKANORI3, Hamano MOMOKO1
1Department of Bioscience and Bioinfomatics, Kyushu Institute of Technology
2Department of Frontier Cardiovascular Science, Graduate School of Medicine, The University of Tokyo Hospital
3Department of Cardiovascular Medicine, Saga University Hospital
Atrial fibrillation (AF) is a common cardiac arrhythmia caused by disorganized electrical signals within the atria. Recent study has reported that several histological features in atrial biopsy have changed and are associated with AF progression. Although individualized treatment based on each patient’s condition is essential, current clinical guidelines lack a standardized framework for histopathological image-based severity classification. In this study, we developed a deep learning-based method that stratifies AF severity into three pathological subtypes using atrial histological images. We also evaluated the clinical utility of the proposed method by exploring the correlation between subtypes and patients’ outcomes.
Hematoxylin and eosin staining images were obtained from atrial biopsy of 533 patients with AF. To capture extensive histological features within the images, we first performed a preprocessing involving image tiling, feature extraction, and integration of RGB-based features. These features were clustered using Kernel K-means, which demonstrated the highest classification performance (Silhouette score: 0.58). Subsequently, the cluster labels derived from this clustering method were used to train a convolutional neural network (CNN) for supervised classification. As a result, CNN model achieved a high accuracy of 0.914. Grad-CAM++ was applied to visualize the discriminative regions contributing to classification, enabling the identification of cluster-specific histological features for each cluster. The heatmap highlighted the cytoplasm of myocytes in Cluster 1, the nuclei of myocytes in Cluster 2, and the intercellular space in Cluster 3, which revealed consistency with the regions of pathologist’s evaluation. Moreover, histological features revealed a marked increase in associated with AF progression in Cluster 3.
Next, we evaluated the clinical relevance of the patients in each cluster. Notably, Cluster 3 exhibited a significantly more patients who experienced recurrence after ablation (Fisher’s exact test: p=7.4×10-3) and composite events (p=6.9×10-3), with odds ratios of 2.53 and 10.4, respectively, compared to Cluster 1. Significant associations were observed with diastolic blood pressure (Kruskal-Wallis test: p=2.6×10-4) and the myocardial injury marker troponin T (p=4.7×10-8). Although the difference was not statistically significant, the prevalence of heart failure with preserved ejection fraction (HFpEF) was highest in Cluster 3, and the proportion of patients with long-standing persistent AF increased progressively from Cluster 1 to Cluster 3.
These findings unveiled the potential of computational pathology diagnosis to identify clinically meaningful subtypes of AF that are not readily discernible through conventional diagnostics. By combining deep learning and histopathological image analysis, this study demonstrated a novel framework for evaluating AF severity and facilitating personalized treatment strategies.