Digital Health -> Digital Health D-PO01 - Featured Poster Session (ID 11) Poster

D-PO01-065 - Prediction Of Shock Impedance In Subcutaneous Implantable Defibrillators Using Chest Radiographs Utilizing A Deep Convolutional Neural Network (ID 68)


Background: Options for evaluating subcutaneous implantable cardioverter-defibrillator (S-ICD) shock efficacy without inducing ventricular arrhythmia are limited. Anatomic characteristics of lead and generator placement are associated with shock efficacy but can only be assessed on post-implant chest imaging. Low shock impedance has also been associated with higher likelihood of successful defibrillation. Strategies to predict shock impedance may allow S-ICD efficacy to be evaluated without DFT testing at implant.
Objective: To develop and validate a neural network classification model for predicting S-ICD shock impedance from chest radiographs.
Methods: We identified 141 patients (631 images) who underwent S-ICD placement with defibrillation threshold testing at 2 centers and obtained chest radiographs (anterior-posterior and posterior-anterior). Radiographs were de-identified and coded based on shock impedance (>70 ohms, < 70 ohms). We split the data into training (80%), validation (10%), and testing (10%) sets. We built a convolutional neural network (CNN, Fig A) using the training images and tested the network using the validation images.
Results: At implant, mean age of S-ICD recipients was 58.4±18.9 years,80% were implanted for primary prevention and 78% had non-ischemic substrates. Example of chest image is shows in fig. B and fig. C. On the training images (505 images, 121 patients), the CNN demonstrated accuracy of 73.1%, sensitivity of 68.1%, and specificity of 73.5% for predicting shock impedance > or < 70 ohms. In testing on 63 images, the CNN achieved an accuracy of 70.1%, specificity of 69.1%, and sensitivity of 72.3%.
Conclusion: This proof of concept study demonstrates a neural network may be able to predict shock impedance based on a frontal chest radiograph. Further data is needed to better train and validate this model, which may provide insight on characteristics associated with shock impedance which have not yet been identified. Additionally, these findings can be translated into a single AP fluoroscopy image and may obviate the need for post-procedure imaging.