Patient-Level Out-of-Fold-Guided Threshold-Aware Heterogeneous Convolutional Neural Network–XGBoost Probability Fusion for Identification of Atrial Fibrillation from PTB-XL

Atrial fibrillation (AF) is a prevalent cardiac rhythm disorder that requires accurate and reliable detection from electrocardiogram (ECG) recordings. This study proposes a patient-level, threshold-aware heterogeneous CNN–XGBoost probability-fusion framework that integrates complementary waveform-based and engineered ECG representations for AF detection using the PTB-XL database. The study included 21,799 twelve-lead ECG recordings, comprising 1,514 AF and 20,285 non-AF recordings. A total of 91 ECG-derived temporal, morphological, statistical, and frequency-domain features were used by the XGBoost branch, while a convolutional neural network (CNN) model learned representations directly from preprocessed twelve-lead ECG waveforms. Patient-level partitioning was used to prevent patient overlap between development and independent test sets. Within the development set, strict five-fold out-of-fold (OOF) predictions were generated for model comparison, fusion-weight selection, and decision-threshold selection, while the independent test set remained completely untouched until final evaluation. CNN and XGBoost probabilities were combined using development-OOF-selected weights of 0.75 and 0.25, respectively, and the decision threshold was fixed at 0.740 before independent testing. On 4,429 independent test ECG recordings, the proposed fusion achieved 98.56% accuracy, 90.29% precision, 89.14% sensitivity, 99.27% specificity, 89.71% F1-score, 94.20% balanced accuracy, 0.889 Matthews correlation coefficient (MCC), 0.889 Cohen’s kappa, 0.9897 ROC-AUC, and 0.9453 PR-AUC, with a Brier score of 0.0134. Compared with the CNN alone, fusion produced a larger observed increase in PR-AUC than in ROC-AUC, while also yielding a lower Brier score. These findings show that integrating complementary waveform-based and engineered ECG representations was associated with improved precision–recall performance under leakage-controlled patient-level evaluation. External validation using independent multicentre and prospective ECG datasets is required to assess generalizability.