System and method for atrial fibrillation detection in non-noise ECG data with the aid of a digital computer
A system and method for atrial fibrillation detection in non-noise ECG data with the aid of a digital computer are provided. Electrocardiography (ECG) features and annotated patterns of the features are maintained in a database, at least some of the patterns associated with atrial fibrillation. A classifier is trained based on the annotated patterns, the classifier implemented by a convolutional neural network. A representation of an ECG signal recorded by one or more ambulatory monitors is received. Noise is detected in the representations and ECG features in the representation falling within each of the non-noise temporal windows are detected. The trained classifier is used to identify patterns of the ECG features. A value indicative of whether portions of the representation are associated the patient experiencing atrial fibrillation is calculated. That one or more of the portions are associated with the patient experiencing atrial fibrillation is determined.
1. A system for atrial fibrillation detection in non-noise ECG data with the aid of a digital computer, comprising:
a database operable to maintain a plurality of electrocardiography (ECG) features and annotated patterns of the features, at least some of the patterns associated with atrial fibrillation;
at least one server interconnected to the database and implementing a convolutional neural network, the convolutional neural network comprising two convolutional hidden layers and two fully-connected hidden layers, the at least one server configured to:
train a classifier based on the annotated patterns in the database;
receive a representation of an ECG signal recorded by one or more cardiac monitors;
identify portions of the representation that comprise noise;
detect a plurality of the ECG features in non-noise portions of the representation;
use the trained classifier to identify patterns of the ECG features within the non-noise portions of the ECG signal;
for each of the portions, calculate a value indicative of whether that portion of the representation is associated the patient experiencing atrial fibrillation;
determine that one or more of the non-noise portions of the ECG signal are associated with the patient experiencing atrial fibrillation using the values for those non-noise portions.
2. A system according to claim 1 , the at least one server further configured to:
test the trained classifier, comprising:
obtain a set of testing data;
classify the testing data using the trained classifier;
obtain classifications of the testing data from a source different than the trained classifier; and
compare the trained classifier classifications to the different source classifications; and
perform further training of the trained classifier following the comparison.
3. A system according to claim 2 , the at least one server configured to:
determine an accuracy of the classifications of the testing data by the trained classifier based upon the comparison; and
compare the accuracy to a threshold, wherein the further training is performed upon the accuracy falling below the threshold.
4. A system according to claim 3 , wherein the testing data comprises a plurality of batches of samples and an average accuracy is determined as the accuracy after the classification of each one of the batches.
5. A system according to claim 1 , wherein the ECG features are detected in non-noise portions of the representation using sliding filters.
6. A system according to claim 1 , the at least one server further configured to:
generate at least one matrix with weights for the patterns, wherein the value is generated using the matrix.
7. A system according to claim 1 , the at least one server further configured to:
generate a further value for each of the non-noise portions based on the patterns that is indicative of the portions being associated with the patient not experiencing atrial fibrillation, wherein the determination for each of the portions is further based on the further value for that portion.
8. A system according to claim 1 , wherein the ECG features comprise one or more of R-waves without clearly discernible P-waves and disorganized P-waves that do not settle into a repeating pattern.
9. A system according to claim 1 , wherein the patterns comprise irregular intervals between successive R-waves.
10. A system according to claim 1 , the at least one server further configured to take an action upon the determination.
11. A method for atrial fibrillation detection in non-noise ECG data with the aid of a digital computer, comprising:
maintaining in a database a plurality of electrocardiography (ECG) features and annotated patterns of the features, at least some of the patterns associated with atrial fibrillation;
training a classifier based on the annotated patterns in the database by at least one server that is interconnected to the database and that implements a convolutional neural network, the convolutional neural network comprising two convolutional hidden layers and two fully-connected hidden layers;
receiving by the at least one server a representation of an ECG signal recorded by one or more cardiac monitors;
identifying by the at least one server portions of the representation that comprise noise;
detecting by the at least one server a plurality of the ECG features in non-noise portions of the representation;
using by the at least one server the trained classifier to identify patterns of the ECG features within the non-noise portions of the ECG signal;
for each of the portions, calculating by the at least one server a value indicative of whether that portion of the representation is associated the patient experiencing atrial fibrillation;
determining by the at least one server that one or more of the non-noise portions of the ECG signal are associated with the patient experiencing atrial fibrillation using the values for those non-noise portions.
12. A method according to claim 11 , further comprising:
testing the trained classifier, comprising:
obtaining a set of testing data;
classifying the testing data using the trained classifier;
obtaining classifications of the testing data from a source different than the trained classifier; and
comparing the trained classifier classifications to the different source classifications; and
performing further training of the trained classifier following the comparison.
13. A method according to claim 12 , further comprising:
determining an accuracy of the classifications of the testing data by the trained classifier based upon the comparison; and
comparing the accuracy to a threshold, wherein the further training is performed upon the accuracy falling below the threshold.
14. A method according to claim 13 , wherein the testing data comprises a plurality of batches of samples and an average accuracy is determined as the accuracy after the classification of each one of the batches.
15. A method according to claim 11 , wherein the ECG features are detected in non-noise portions of the representation using sliding filters.
16. A method according to claim 11 , further comprising:
generating at least one matrix with weights for the patterns, wherein the value is generated using the matrix.
17. A method according to claim 11 , the at least one server further configured to:
generating a further value for each of the non-noise portions based on the patterns that is indicative of the portions being associated with the patient not experiencing atrial fibrillation, wherein the determination for each of the portions is further based on the further value for that portion.
18. A method according to claim 11 , wherein the ECG features comprise one or more of R-waves without clearly discernible P-waves and disorganized P-waves that do not settle into a repeating pattern.
19. A method according to claim 11 , wherein the patterns comprise irregular intervals between successive R-waves.
20. A method according to claim 11 , further comprising taking an action upon the determination.