Development and validation of explainable deep learning models for classification of atrial fibrillation subtypes using cardiac computed tomography

Background Although cardiac computed tomography (CT) provides detailed anatomical information on the left atrial (LA), few studies have examined whether it can distinguish paroxysmal atrial fibrillation (PAF) from persistent atrial fibrillation (PerAF) based on structural features in an interpretable manner. Objective To develop a convolutional neural network (CNN) model trained on LA morphology derived from cardiac CT for classifying atrial fibrillation (AF) subtypes and to identify spatial remodeling patterns associated with PerAF to enhance understanding of AF progression. Methods We developed 3 types of 3-dimensional CNN to classify AF subtypes using cardiac CT-derived LA morphology. A total of 269 patients were used for model development with stratified 10-fold cross-validation. External validation was conducted in 151 independent patients. CNN performance was compared with LA volume and LA volume index from echocardiography and CT. We used gradient-weighted class activation mapping to identify regional remodeling patterns associated with predictions. Results Among the 3-dimensional-CNN, the 3D-DenseNet201 model achieved the highest performance in internal validation (area under the receiver operating characteristic curve 0.81 ± 0.08; accuracy 77.0 ± 6.2%) and maintained consistent accuracy in external validation (area under the receiver operating characteristic curve 0.81 ± 0.01; accuracy 76.7 ± 1.6%). gradient-weighted class activation mapping revealed that PerAF classification was primarily driven by activation in the anterosuperior LA wall (72.8%), right superior pulmonary vein antrum (49.4%), and septum (44.3%). The posterior wall showed minimal activation. CNN outperformed echocardiographic or CT-derived volume metrics. Conclusion The 3D-DenseNet201 model accurately classified AF subtypes and localized structural remodeling patterns relevant to PerAF. These findings highlight the potential of deep learning to improve the mechanistic understanding of AF progression.