A CONVOLUTIONAL NEURAL NETWORK FOR AUTOMATIC DISCRIMINATION OF PAUSE EPISODES DETECTED BY AN INSERTABLE CARDIAC MONITOR

Abstract

BackgroundInsertable cardiac monitors (ICMs) are widely used to detect pause events. False detections may create a considerable additional review burden for clinics. Due to computational constraints, device-based algorithms have not been able to eliminate these false detections. Convolutional Neural Networks (CNNs), with automatic feature extraction, have demonstrated decent performance for ECG classification in recent studies. A CNN can take an input and assign importance (learnable weights and biases) to aspects of the input that are critical to providing an accurate prediction. These self-extracted features are learned during training with labeled input.ObjectiveDevelop and test a new CNN model for the automatic classification of pause episodes detected by ICMs.MethodsEach ICM pause electrogram (EGM) is input into a compact 1D CNN with 5 layers to output an adjudication prediction. The CNN model was trained and tested using expert adjudicated pause EGMs detected by Confirm Rx™ and Jot Dx™ ICMs. Data augmentation techniques, such as noise addition and signal inversion, were used in training. EGMs in the test data set were randomly selected from ICMs that were not used in training. Sensitivity, specificity, and positive predictive value (PPV) metrics were used to assess the performance of the CNN model.ResultsThe test data set from 750 ICM devices included 1,982 pause EGMs, of which 752 EGMs were adjudicated as true positives. The CNN model achieved a sensitivity of 98.9% and specificity of 94.6%. It also reduced the number of EGMs for review by 59.0% to 811 EGMs. Of the 482 devices with false positive episodes, 429 (89.0%) became FP free after CNN classification. The PPV of pause EGMs improved from 37.9% to 91.7%.Conclusion BackgroundInsertable cardiac monitors (ICMs) are widely used to detect pause events. False detections may create a considerable additional review burden for clinics. Due to computational constraints, device-based algorithms have not been able to eliminate these false detections. Convolutional Neural Networks (CNNs), with automatic feature extraction, have demonstrated decent performance for ECG classification in recent studies. A CNN can take an input and assign importance (learnable weights and biases) to aspects of the input that are critical to providing an accurate prediction. These self-extracted features are learned during training with labeled input. Insertable cardiac monitors (ICMs) are widely used to detect pause events. False detections may create a considerable additional review burden for clinics. Due to computational constraints, device-based algorithms have not been able to eliminate these false detections. Convolutional Neural Networks (CNNs), with automatic feature extraction, have demonstrated decent performance for ECG classification in recent studies. A CNN can take an input and assign importance (learnable weights and biases) to aspects of the input that are critical to providing an accurate prediction. These self-extracted features are learned during training with labeled input. ObjectiveDevelop and test a new CNN model for the automatic classification of pause episodes detected by ICMs. Develop and test a new CNN model for the automatic classification of pause episodes detected by ICMs. MethodsEach ICM pause electrogram (EGM) is input into a compact 1D CNN with 5 layers to output an adjudication prediction. The CNN model was trained and tested using expert adjudicated pause EGMs detected by Confirm Rx™ and Jot Dx™ ICMs. Data augmentation techniques, such as noise addition and signal inversion, were used in training. EGMs in the test data set were randomly selected from ICMs that were not used in training. Sensitivity, specificity, and positive predictive value (PPV) metrics were used to assess the performance of the CNN model. Each ICM pause electrogram (EGM) is input into a compact 1D CNN with 5 layers to output an adjudication prediction. The CNN model was trained and tested using expert adjudicated pause EGMs detected by Confirm Rx™ and Jot Dx™ ICMs. Data augmentation techniques, such as noise addition and signal inversion, were used in training. EGMs in the test data set were randomly selected from ICMs that were not used in training. Sensitivity, specificity, and positive predictive value (PPV) metrics were used to assess the performance of the CNN model. ResultsThe test data set from 750 ICM devices included 1,982 pause EGMs, of which 752 EGMs were adjudicated as true positives. The CNN model achieved a sensitivity of 98.9% and specificity of 94.6%. It also reduced the number of EGMs for review by 59.0% to 811 EGMs. Of the 482 devices with false positive episodes, 429 (89.0%) became FP free after CNN classification. The PPV of pause EGMs improved from 37.9% to 91.7%. The test data set from 750 ICM devices included 1,982 pause EGMs, of which 752 EGMs were adjudicated as true positives. The CNN model achieved a sensitivity of 98.9% and specificity of 94.6%. It also reduced the number of EGMs for review by 59.0% to 811 EGMs. Of the 482 devices with false positive episodes, 429 (89.0%) became FP free after CNN classification. The PPV of pause EGMs improved from 37.9% to 91.7%. Conclusion