End-to-end joint Margin aware BiLSTM-dSVM model for atrial fibrillation detection based on knowledge-augmented deep learning on knowledge-augmented deep learning

To avoid stroke, atrial fibrillation (AF) needs to be detected as early as possible. Nevertheless, traditional deep learning systems often are unsuccessful with the domain shift and signal degradation inherent in wearable sensor measurements. To achieve the purpose of going beyond probabilistic classification, We present a hybrid convolutional neural network–bidirectional long ahort-term memory–differentiable support vector machine (CNN–BiLSTM–dSVM) framework that integrates automated morphological feature generation with geometric margin maximization. The main methodological innovation is the (dSVM) head, which is trained on a aquared hinge loss objective, instead of conventional Soft max layers. To isolate structural P-QRS-T variations and remove motion artifacts and baseline drift, this architecture uses a CNN backbone. These morphological signs are used with knowledge-augmented clinical data to represent long-term temporal rhythms and are processed by a Bi LSTM encoder. Unlike present models, our method removes overfitting to sensor-induced noise, by imposing structural risk minimization, and creating an operative loss-friendly dead zone. Experimental analysis of the proposed system was conducted on dual-source ECG datasets.The proposed framework was evaluated using dual-source ECG datasets, including the MIT-BIH atrial fibrillation database, the PhysioNet CinC 2017 dataset, and a noisy wearable ECG dataset collected from real-world sensing environments. The architecture combined CNN-based morphological feature extraction, BiLSTM temporal representation learning, and a (dSVM) optimization strategy to improve atrial fibrillation classification robustness under heterogeneous ECG acquisition conditions. Experimental results demonstrated that the proposed framework achieved an overall classification accuracy of 99.0%, recall of 98.8%, Macro F1-score of 98.6%, and AUC of 0.995. Comparative evaluation against mainstream baseline architectures, including conventional CNN, LSTM, and Transformer-based models, confirmed superior convergence stability, generalization capability, and classification robustness, particularly under noisy wearable ECG conditions. (20,000 noisy wearable recordings and 6,000 high-fidelity clinical signals). The datasets were partitioned into training, validation, and testing subsets using a strict patient-level splitting protocol to avoid data leakage. This gave system a near-perfect overall accuracy of 99.5% and a recall of 1.00. The proposed hybrid model restored this performance in 400 Hz noiseful mobile signals, but standalone CNNs exhibited constraints in terms of dealing with class imbalances (Macro F1: 0.593). The sensor-invariant representation of our Margin-aware inductive bias provided better convergence kinetics and structural stability than high-complexity 1D-transformer and ECG-BERT baselines as found in statistical studies. In this work, we established a benchmark of computationally efficient, theoretically-grounded cardiac surveillance in real-world mHealth.