Recurrent neural network architecture based classification of atrial fibrillation using single lead ECG

Patent metadata

Field	Value
Publication number	US-11571162-B2
Application number	US-202016827812-A
Country	US
Kind code	B2
Filing date	Mar 24, 2020
Priority date	May 9, 2019
Publication date	Feb 7, 2023
Grant date	Feb 7, 2023

Abstract

Conventionally, Atrial Fibrillation (AF) has been detected using atrial analyses which is vulnerable to background noise. Again there is a dependency on statistical features which are extracted from R-R intervals of long ECG recordings. The present disclosure addresses AF detection from single lead short ECG recordings of less than one minute wherein automatic detection of P-R and P-Q intervals is difficult, which introduces error in feature computing from the segregated intervals and compromises the performance of the classifier. In the present disclosure, a Recurrent Neural Network (RNN) based architecture comprising two Long Short Term Memory (LSTM) networks is provided for temporal analysis of R-R intervals and P wave regions in an ECG signal respectively. Output sates of the two LSTM networks are merged at a dense layer along with a set of hand-crafted statistical features to create a composite feature set for classification of the AF.

First claim

What is claimed is:

1. A processor implemented method for classification of atrial fibrillation (AF), the method comprising the steps of:
acquiring, by one or more hardware processors, a single lead electrocardiogram (ECG) recorded for a first pre-defined time period being less than a minute;
obtaining, by the one or more hardware processors, a first time series being an R-R interval time series based on R peaks in the acquired ECG;
identifying, by the one or more hardware processors, a region having a second pre-defined time period before each of the R peaks to form a second time series, wherein the second time series is a region corresponding to P wave time series and the second pre-defined time period is in the range of 120-200 milliseconds;
inputting, by the one or more hardware processors, the first time series and the second time series independently to a pair of Long short-term memory (LSTM) networks, wherein a Bidirectional LSTM (BiLSTM) network and an LSTM network constitute the pair, and wherein the step of inputting the first time series and the second time series independently to the pair of LSTM networks is followed by:
performing a temporal analyses of the R-R intervals using the BiLSTM network to capture irregular R-R intervals; and
performing a temporal analyses of atrial activities using the LSTM network, wherein the atrial activities include absence of P waves or presence of f-waves before a QRS complex in the acquired ECG;

merging, by the one or more hardware processors, output states of the LSTM network associated with the first time series and the second time series along with a pre-defined set of handcrafted statistical features computed from the acquired ECG to create a composite feature set for classification of the AF; and
classifying the AF in the acquired ECG based on the composite feature set using a classifier.

2. The processor implemented method of claim 1 , wherein the first pre-defined time period is 33 seconds.

3. The processor implemented method of claim 1 , wherein the second pre-defined time period is 200 milliseconds.

4. The processor implemented method of claim 1 , wherein the region having the second pre-defined time period represents a window before the QRS complex in the acquired ECG where the P wave is located and the second time series comprises a plurality of windows on time axis.

5. The processor implemented method of claim 1 , wherein the cardinality of the pre-defined set of handcrafted statistical features is 20.

6. A system for classification of atrial fibrillation, the system comprising:
one or more hardware processors;
one or more data storage devices operatively coupled to the one or more hardware processors and configured to store instructions configured for execution by the one or more hardware processors to:
acquire a single lead electrocardiogram (ECG) recorded for a first pre-defined time period being less than a minute;
obtain a first time series being an R-R interval time series based on R peaks in the acquired ECG;
identify a region having a second pre-defined time period before each of the R peaks to form a second time series, wherein the second time series is a region corresponding to P wave time series and the second pre-defined time period is in the range of 120-200 milliseconds;
input the first time series and the second time series independently to a pair of Long short-term memory (LSTM) networks; and
merge output states of the LSTM network associated with the first time series and the second time series along with a pre-defined set of handcrafted statistical features computed from the acquired ECG to create a composite feature set for classification of the AF;

the pair of the LSTM networks, wherein a Bidirectional LSTM (BiLSTM) network and an LSTM network constitute the pair, and wherein the BiLSTM is configured to receive the first time series and perform a temporal analyses of the R-R intervals to capture irregular R-R intervals and the LSTM network is configured to receive the second time series and perform a temporal analyses of atrial activities, wherein the atrial activities include absence of P waves or presence of f-waves before a QRS complex in the acquired ECG; and
a classifier comprising a plurality of full connected layers and a softmax function, wherein the classifier is configured to classify the AF in the acquired ECG based on the composite feature set.

7. The system of claim 6 , wherein the first pre-defined time period is 33 seconds.

8. The system of claim 6 , wherein the second pre-defined time period is 200 milliseconds.

9. The system of claim 6 , wherein the region having the second pre-defined time period represents a window before a QRS complex in the acquired ECG where the P wave is located and the second time series comprises a plurality of windows on time axis.

10. A computer program product comprising a non-transitory computer readable medium having a computer readable program embodied therein, wherein the computer readable program, when executed on a computing device, causes the computing device to:
acquire a single lead electrocardiogram (ECG) recorded for a first pre-defined time period being less than a minute;
obtain a first time series being an R-R interval time series based on R peaks in the acquired ECG;
identify a region having a second pre-defined time period before each of the R peaks to form a second time series, wherein the second time series is a region corresponding to P wave time series and the second pre-defined time period is in the range of 120-200 milliseconds;
input the first time series and the second time series independently to a pair of Long short-term memory (LSTM) networks, wherein a Bidirectional LSTM (BiLSTM) network and an LSTM network constitute the pair, and wherein the step of inputting the first time series and the second time series independently to the pair of LSTM networks is followed by:
performing a temporal analyses of the R-R intervals using the BiLSTM network to capture irregular R-R intervals; and
performing a temporal analyses of atrial activities using the LSTM network, wherein the atrial activities include absence of P waves or presence of f-waves before a QRS complex in the acquired ECG;

merge output states of the LSTM network associated with the first time series and the second time series along with a pre-defined set of handcrafted statistical features computed from the acquired ECG to create a composite feature set for classification of the AF; and
classify the AF in the acquired ECG based on the composite feature set using a classifier.

Assignees

• Tata Consultancy Services Ltd

Inventors

• Banerjee Rohan
• Ghose Avik
• Khandelwal Sundeep

Classifications

• G06N3/0442

  characterised by memory or gating, e.g. long short-term memory [LSTM] or gated recurrent units [GRU] · CPC title

• G06N3/09

  Supervised learning · CPC title

• G06N3/045

  Combinations of networks · CPC title

• A61B5/7264 Primary

  Classification of physiological signals or data, e.g. using neural networks, statistical classifiers, expert systems or fuzzy systems · CPC title

• A61B5/361 Primary

  Detecting fibrillation · CPC title