gms | German Medical Science

Automated feature extraction using deep learning significantly enhances pattern recognition capabilities beyond traditional manual feature engineering methods. This is exemplified in the convolutional layers of the ResNet architecture, derived from the visual cortex, which autonomously learns relevant patterns (features) from data. Despite the high sensitivity and specificity of commercially available automated external defibrillators (AEDs), approximately 10% of shockable conditions are still missed, highlighting a need for improvement ([1], p. 23).

This study utilizes residual neural networks (ResNets) to classify ECG data for shockable heart rhythms, aiming to improve survival rates for sudden cardiac arrest patients. We employed four renowned public databases: the MIT-BIH Arrhythmia Database (MITDB), Creighton University Ventricular Tachycardia Database (CUDB), MIT-BIH Ventricular Arrhythmia Database (VFDB), and the American Heart Association Database (AHADB), which collectively contribute to advancing algorithm development [2], [3], [4].

A total of 125 Holter recordings were processed, creating a balanced dataset of 60,340 segments, 45% of which represented shockable conditions. The dataset was augmented from segments around and along the borders of shockable episodes to address the notable imbalance where non-shockable labels predominated at 86.86%, thereby enhancing the model's exposure to critical transition points between conditions. The ECG signals underwent noise reduction, wavelet transformation, and segmentation into 3-second intervals. Specific preprocessing techniques included uniform resampling at 250 Hz [2], [4], application of a high-pass filter with a 1 Hz cutoff, and a second-order Butterworth low-pass filter with a 30 Hz cutoff to reduce muscle noise and powerline interference.

Our ResNet model, optimized from 70 iterations, was rigorously tested through leave-one-subject-out cross-validation on blocks of three consecutive segments, or 9-second blocks, to enhance shock advisement reliability and reduce false positives. The architecture included three residual stages with three blocks each and concluded with four dense layers, achieving an accuracy of 99.64%, sensitivity of 99.55%, and specificity of 99.48% in detecting shockable conditions.

Further validation involved manual annotation by three cardiologists to assess misclassifications, revealing variability in annotations regarding shockable episode onset and termination. This insight points to potential inconsistencies in training data labels that could affect learning and test performance. To our knowledge, the mentioned performance values and the analysis of misclassified segments complement and extend the work of comparable studies [2], [3], [4], [5].

The findings confirm that integrating automated feature detection with neural networks can improve ECG data analysis, extending beyond shockable cardiac arrhythmias to other biosignal analyses. Future work will focus on broader database verification and potential mobile application implementations.

The authors declare that they have no competing interests.

The authors declare that an ethics committee vote is not required.