Multi-layer perceptron neural network-based algorithm for automatic external defibrillator

Traditional rule-based automatic external defibrillator (AED) algorithms encounter difficulties in accurately classifying electrocardiograms (ECGs) in the borderline area. To overcome this limitation, we have developed an artificial intelligence (AI)-based AED algorithm.

We utilized open-source databases, hospital ECG databases, and AED databases, acquiring a total of 1,551,089 segments. For the learning phase, 14,000 shockable (25%) and 42,000 non-shockable segments (75%) were allocated. Deep learning was conducted using Multi-layer Perceptron Neural Networks (MLPNN) and fully connected neural networks (FCNN). ECG-feature extractions included threshold crossing sample count (TCSC), ventricular fibrillation-filter leak (VF-leak), phase space reconstruction (PSR), QRS complex features (bCP, bWT, bW), heart rate (bpm), standard deviation of RR interval (RR-std), and standard deviation of peak QRS amplitude (Peak-std). The MLPNN configuration included four hidden layers (see Figures 1 and 2). Activation functions such as hyperbolic tangent, softmax, and sigmoid were applied, with Binary Cross-Entropy serving as the loss function. To prevent overfitting, a 10-fold cross-validation approach was employed.

Testing on database segments (n = 1,468,879) demonstrated an accuracy of 99.79%, a precision of 98.66%, a recall of 93.67%, and an F1 score of 96.10%. Sensitivity for ventricular fibrillation was 98.7%, ventricular tachycardia sensitivity was 81.0%, normal sinus rhythm specificity was 100%, asystole specificity was 100%, and non-shockable rhythm specificity was 99.8%. Overall sensitivity was 93.67%, specificity was 99.96%, and the AUC was 96.82%.

The MLPNN-based AED algorithm, incorporating ECG-feature extractions (TCSC, VF-leak, PSR, bCP, bWT, bW, bpm, RR-std, and Peak-std), exhibited high accuracy, good precision, and reproducibility