Methods and Systems for Determining Abnormal Cardiac Activity

Abstract

The systems and methods can accurately and efficiently determine abnormal cardiac activity from motion data and/or cardiac data using techniques that can be used for long-term monitoring of a patient. In some embodiments, the method for using machine learning to determine abnormal cardiac activity may include receiving one or more may include applying a trained deep learning architecture to each tensor of the one or more periods of time to classify each window and/or each period into one or more classes using at least the one or more signal quality indices for the cardiac data and the motion data and cardiovascular features. The deep learning architecture may include a convolutional neural network, a bidirectional recurrent neural network, and an attention network. The one or more classes may include abnormal cardiac activity and normal cardiac activity.

1 . A computer-implemented method for using machine learning to determine abnormal cardiac activity of a subject, the method comprising:
 receiving one or more periods of time of cardiac data and motion data for a subject, each period of time including more than one window of the cardiac data and the motion data;   determining one or more signal quality indices for each window of the cardiac data and the motion data of the one or more periods of time;   extracting one or more cardiovascular features for each period of time using at least the cardiac data, the motion data, and the one or more signal quality indices for the cardiac data and the motion data;   applying a tensor transform to the cardiac data and/or the motion data to generate a tensor for each window of the one or more periods of time;   applying a trained deep learning architecture to each tensor of the one or more periods of time to classify each window and/or each period into one or more classes using at least the one or more signal quality indices for the cardiac data and the motion data and cardiovascular features,   the deep learning architecture including a convolutional neural network, a bidirectional recurrent neural network, and an attention network,   the one or more classes including abnormal cardiac activity and normal cardiac activity; and   generating a report including a classification of cardiac activity of the subject for the one or more periods based on the one or more classes.   
     
     
         2 . The method according to  claim 1 , further comprising:
 receiving subject contextual information for the subject, the subject contextual information including medical history and demographic information;   wherein the extracting uses one or more subject information features related to the subject contextual information to extract one or more cardiovascular features for each period of time, and the trained deep learning architecture uses the one or more subject information features to classify the cardiac activity for each window of the period.   
     
     
         3 . The method according to  claim 1 , wherein the tensor transform is applied to the cardiac data and the motion data for each window. 
     
     
         4 . The method according to  claim 3 , further comprising:
 determining a quality channel for each window based on the one or more signal quality indices for the cardiac data and the motion data, the quality channel corresponding to a channel in each window having the one more quality indices that is higher than remaining channels in each channel.   
     
     
         5 . The method according to  claim 1 , wherein the applying the deep learning architecture includes:
 encoding each tensor for each window of the one or more periods using the deep convolutional network into one or more deep learning features associated with cardiac activity;   applying the bidirectional recurrent network to determine a probability that each window of the one or more periods belongs to a class of the one or more classes, the bidirectional recurrent network using the one or more deep learning features, the one more signal quality indices for the cardiac data and/or motion data, and/or one or more cardiovascular features to classify each window of the one or more periods; and   determining the classification of cardiac activity for each window of the one or more periods and/or each period by applying the attention network to the probability for each window of the one or more periods.   
     
     
         6 . The method according to  claim 5 , wherein the attention network determines a score for each window and/or each period, the score representing the classification of cardiac activity. 
     
     
         7 . The method according to any of  claim 6 , wherein when the classification of cardiac activity includes abnormal cardiac activity, a window of each period having a highest score represents the window including the abnormal cardiac activity. 
     
     
         8 . A non-transitory computer-readable storage medium storing instructions for using machine learning to determine abnormal cardiac activity of a subject, the instructions comprising
 receiving one or more periods of time of cardiac data and motion data for a subject, each period of time including more than one window of the cardiac data and the motion data;   determining one or more signal quality indices for each window of the cardiac data and the motion data of the one or more periods of time;   extracting one or more cardiovascular features for each period of time using at least the cardiac data, the motion data, and the one or more signal quality indices for the cardiac data and the motion data;   applying a tensor transform to the cardiac data and/or the motion data to generate a tensor for each window of the one or more periods of time;   applying a trained deep learning architecture to each tensor of the one or more periods of time to classify each window and/or each period into one or more classes using at least the one or more signal quality indices for the cardiac data and the motion data and cardiovascular features,   the deep learning architecture including a convolutional neural network, a bidirectional recurrent neural network, and an attention network,   the one or more classes including abnormal cardiac activity and normal cardiac activity; and   generating a report including a classification of cardiac activity of the subject for the one or more periods based on the one or more classes.   
     
     
         9 . The medium according to  claim 8 , the instructions further comprising:
 receiving subject contextual information for the subject, the subject contextual information including medical history and demographic information;   wherein the extracting uses one or more subject information features related to the subject contextual information to extract one or more cardiovascular features for each period of time, and the trained deep learning architecture uses the one or more subject information features to classify the cardiac activity for each window of the period.   
     
     
         10 . The medium according to  claim 8 , wherein the tensor transform is applied to the cardiac data and the motion data for each window. 
     
     
         11 . The medium according to  claim 10 , further comprising:
 determining a quality channel for each window based on the one or more signal quality indices for the cardiac data and the motion data, the quality channel corresponding to a channel in each window having the one more quality indices that is higher than remaining channels in each channel.   
     
     
         12 . The medium according to  claim 8 , wherein the applying the deep learning architecture includes:
 encoding each tensor for each window of the one or more periods using the deep convolutional network into one or more deep learning features associated with cardiac activity;   applying the bidirectional recurrent network to determine a probability that each window of the one or more periods belongs to a class of the one or more classes, the bidirectional recurrent   network using the one or more deep learning features, the one more signal quality indices for the cardiac data and/or motion data, and/or one or more cardiovascular features to classify each window of the one or more periods; and   determining the classification of cardiac activity for each window of the one or more periods and/or each period by applying the attention network to the probability for each window of the one or more periods.   
     
     
         13 . The medium according to  claim 12 , wherein:
 the attention network determines a score for each window and/or each period, the score representing the classification of cardiac activity; and   when the classification of cardiac activity includes abnormal cardiac activity, a window of each period having a highest score represents the window including the abnormal cardiac activity.   
     
     
         14 . A system for using machine learning to determine abnormal cardiac activity of a subject, comprising:
 a memory; and   one or more processors, wherein the one or more processors is configured to cause:
 receiving one or more periods of time of cardiac data and motion data for a subject, each period of time including more than one window of the cardiac data and the motion data; 
 determining one or more signal quality indices for each window of the cardiac data and the motion data of the one or more periods of time; 
 extracting one or more cardiovascular features for each period of time using at least the cardiac data, the motion data, and the one or more signal quality indices for the cardiac data and the motion data; 
 applying a tensor transform to the cardiac data and/or the motion data to generate a tensor for each window of the one or more periods of time; 
 applying a trained deep learning architecture to each tensor of the one or more periods of time to classify each window and/or each period into one or more classes using at least the one or more signal quality indices for the cardiac data and the motion data and cardiovascular features, 
 the deep learning architecture including a convolutional neural network, a bidirectional recurrent neural network, and an attention network, 
 the one or more classes including abnormal cardiac activity and normal cardiac activity; and 
 generating a report including a classification of cardiac activity of the subject for the one or more periods based on the one or more classes. 
   
     
     
         15 . The system according to  claim 14 , wherein the processor is further configured to cause:
 receiving subject contextual information for the subject, the subject contextual information including medical history and demographic information;   wherein the extracting uses one or more subject information features related to the subject contextual information to extract one or more cardiovascular features for each period of time, and the trained deep learning architecture uses the one or more subject information features to classify the cardiac activity for each window of the period.   
     
     
         16 . The system according to  claim 14 , wherein the tensor transform is applied to the cardiac data and the motion data for each window. 
     
     
         17 . The system according to  claim 16 , further comprising:
 determining a quality channel for each window based on the one or more signal quality indices for the cardiac data and the motion data, the quality channel corresponding to a channel in each window having the one more quality indices that is higher than remaining channels in each channel.   
     
     
         18 . The system according to  claim 14 , wherein the applying the deep learning architecture includes:
 encoding each tensor for each window of the one or more periods using the deep convolutional network into one or more deep learning features associated with cardiac activity;   applying the bidirectional recurrent network to determine a probability that each window of the one or more periods belongs to a class of the one or more classes, the bidirectional recurrent network using the one or more deep learning features, the one more signal quality indices for the cardiac data and/or motion data, and/or one or more cardiovascular features to classify each window of the one or more periods; and   determining the classification of cardiac activity for each window of the one or more periods and/or each period by applying the attention network to the probability for each window of the one or more periods.   
     
     
         19 . The system according to  claim 18 , wherein the attention network determines a score for each window and/or each period, the score representing the classification of cardiac activity. 
     
     
         20 . The system according to  claim 19 , wherein when the classification of cardiac activity includes abnormal cardiac activity, a window of each period having a highest score represents the window including the abnormal cardiac activity.