Purpose: Deep learning is a subfield of machine learning techniques with the ability to automatically learn features from data. Typical deep architectures use convolutional neural networks (CNN). The purpose of this study was to design defibrillator shock advice algorithms based on CNNs using short analysis segments.

Materials and methods: Electronic files were gathered between 2013 and 2015 from the defibrillators (Lifepak 1000/500, Stryker) used to treat out-of-hospital cardiac arrest patients by the basic life support ambulances of the Basque Country. Six clinicians reviewed the electrocardiogram (ECG) of the rhythm analysis intervals, their consensus shock/no-shock decisions were used as ground truth. The analysis intervals were divided into nonoverlapping ECG segments of 2, 3 and 4-seconds. Data were partitioned patient-wise into training (70%) to develop the algorithm, and test (30%) to report the results. A deep learning shock/no-shock algorithm was designed using 2/3 CNN layers and two dense layers. The process was repeated 100 times to statistically characterize the sensitivity (Se), specificity (Sp), and accuracy (Acc) of the shock advice algorithm.

Results: 4216 analyses (852 patients) were reviewed. The consensus was shock for 498 (181) analyses and no-shock for 3718 (790) analyses, with inter-rater Fleissʼ kappa score κ=0.93 (excluding systematic errors). The mean (90% confidence interval) Se, Sp and Acc of the deep learning algorithms were: 96.1 (93.3–98.4), 99.1 (98.4–99.5) and 98.5 (97.9–99.0) for 4-s segments; 95.7 (92.7– 98.1), 99.0 (98.3–99.5) and 98.4 (97.6–99.0) for 3-s segments; and 93.9 (89.3– 97.3), 99.0 (98.4–99.4) and 98.0 (97.3–98.6) for 2-s segments. The minimum 90% Se and 95% Sp recommended by the American Heart Association (AHA) were met for all segment lengths.

Conclusions: Deep learning shock advice algorithms were demonstrated to be compliant with AHA performance using ECG segment lengths as short as 2- seconds.

8 - Deep learning approach for a shock advise algorithm using short electrocardiogram analysis intervals