JMIR Publications

ABSTRACT

Automated detection of shockable cardiac rhythms plays a critical role in improving patient outcomes during sudden cardiac arrest. Conventional automated external defibrillators (AEDs) rely on handcrafted feature engineering, limiting adaptability. This study explores a deep learning-based approach combining Residual Neural Networks (ResNets), ECG spectrogram representations, and block-based evaluation to enhance shockable rhythm detection. Additionally, expert reannotation of misclassified cases refines model performance and highlights inconsistencies in database labels, which can pose challenges for short-segment-based learning approaches. Four public arrhythmia databases (MIT-BIH Arrhythmia, Creighton Ventricular Tachycardia, MIT-BIH Ventricular Arrhythmia, and American Heart Association) were used to develop and optimize the ResNet32 model. A balanced dataset of 60,340 augmented segments was generated for training. The model classifies ECG rhythms using blocks of three consecutive 3-second segments, providing a shock decision over a 9-second window. Performance was evaluated via leave-one-subject-out cross-validation on the original dataset, consisting of 19,802 blocks (2,495 shockable). The model achieved 99.68% accuracy, 99.63% sensitivity, and 99.69% specificity. After expert cardiologist review of misclassified blocks, where 73% of annotations differed from database labels, performance improved to 99.92%, 99.76%, and 99.87%, respectively. The block-based approach reduces false positives from transient rhythms, aligns with clinical assessment practices, and mitigates annotation inconsistencies. While introducing a brief detection delay (up to 9 seconds), it ensures decision reliability and mirrors real-world AED use, where rhythm persistence is evaluated before delivering a shock. Future applications include real-world mobile deployment, generating standardized ECG databases, and detecting other cardiac conditions such as QT prolongation and electrolyte imbalances.