Adaptive Multiscale Dilated Bi-LSTM achieved 90% training accuracy, 88% validation accuracy, 92.91% specificity, and 70.62% sensitivity for ECG anomaly detection on the PTB-XL dataset, enhancing specificity while reducing false positives.
The Adaptive Multiscale Dilated Bi-LSTM network demonstrates high specificity and robust accuracy for automated ECG anomaly detection, offering a potential tool for continuous cardiac monitoring despite challenges with subtle abnormalities.
Effect estimate: Training accuracy 90%, validation accuracy 88%, specificity 92.91%, precision 77.27%, F1 score 0.7379, sensitivity 70.62%, AUC 0.67 (Normal) and 0.57 (Pathological)
Physical activity is very strongly associated with cardiovascular as well as overall health. A plethora of research has evidenced a decline in all-cause mortality rates and a diminished prevalence of cardiovascular ailments, malignancies, and metabolic disorders among individuals whose lifestyle had regular physical exercise. In the context of these beneficial outcomes, the phenomenon of sudden cardiac death (SCD) continues to present a distressing reality, even in seemingly fit athletes competing at elite standards. Sudden Cardiac Death (SCD) is characterized as an unforeseen demise attributable to cardiac causes, and currently, there exists no established medical protocols capable of averting its occurrence. The most effective preventative measure is to consistently monitor vital signs and detect irregularities to facilitate prompt intervention. In this study, a new Adaptive Multiscale Dilated Bidirectional Long Short-Term Memory (Bi-LSTM) network for ECG Anomaly Classification using PTB-XL dataset (PhysioNet 2022) has been developed. This network leverages multiscale dilated convolution to efficiently capture long-range dependencies while preserving fine-grained temporal features. The proposed model demonstrated robust learning capabilities, achieving a training accuracy of 90% and stabilizing at a validation accuracy of 88% through the implementation of mixed precision training and early stopping at 40 epochs. Evaluation metrics highlight the model’s reliability for clinical screening, evidenced by a high specificity of 92.91% and a precision of 77.27% (F1 score: 0.7379), effectively minimizing false positives. However, discrepancies in AUC scores (0.67 for Normal vs. 0.57 for Pathological) and a moderate sensitivity of 70.62% indicate challenges in distinguishing subtle abnormalities, suggesting that future iterations could benefit from targeted data augmentation to address class imbalances.
Neelanaryanan et al. (Sun,) conducted a other in Patients of all ages with diverse cardiac conditions undergoing 12-lead ECG monitoring from the PTB-XL dataset including normal ECG and multiple cardiac pathologies (n=21,837). Adaptive Multiscale Dilated Bidirectional Long Short-Term Memory (Bi-LSTM) network for ECG anomaly classification vs. Not explicitly stated; evaluation against baseline or standard deep learning models implied was evaluated on ECG anomaly detection and classification accuracy, specificity, sensitivity, precision, F1-score, and AUC for multiple cardiac abnormalities (Training accuracy 90%, validation accuracy 88%, specificity 92.91%, precision 77.27%, F1 score 0.7379, sensitivity 70.62%, AUC 0.67 (Normal) and 0.57 (Pathological)). Adaptive Multiscale Dilated Bi-LSTM achieved 90% training accuracy, 88% validation accuracy, 92.91% specificity, and 70.62% sensitivity for ECG anomaly detection on the PTB-XL dataset, enhancing specificity while reducing false positives.