Neural-network based electrocardiogramy monitor and associated methods
Systems and methods detect cortical arousal events from a single time-varying ECG signal that is obtained via single-lead ECG. A pre-trained deep neural network transforms the ECG signal into a sequence of cortical-arousal probabilities. The deep neural network includes an inception module, a residual neural network, and a long short-term memory neural network to identify structure in the ECG signal that distinguishes periods of cortical arousal from periods without cortical arousal.
1. An electrocardiography monitor, comprising:
a processor;
a memory communicably coupled with the processor, the memory storing a deep neural network comprising:
an inception module comprising a plurality of convolutional neural networks (CNNs);
a residual neural network configured to downsample by a downsampling stride that is greater than one;
a feedback path connecting an output of the residual neural network to an input of the residual neural network; and
a long short-term memory (LSTM) neural network; and
an electronic display communicably coupled with the processor;
the memory further storing machine-readable instructions that, when executed by the processor, control the ECG monitor to:
filter a sequence of ECG values with the inception module such that the plurality of CNNs process the sequence of ECG values in parallel, each of the plurality of CNNs outputting a respective one of a plurality of feature maps;
iteratively downsample, with the residual neural network and via the feedback path, the plurality of feature maps to generate a respective plurality of downsampled feature maps;
calculate a sequence of cortical-arousal probabilities by feeding the plurality of downsampled feature maps into the LSTM neural network;
compare each of the sequence of cortical-arousal probabilities to a threshold; and
display, on the electronic display, an indication in response to one or more of the sequence of cortical-arousal probabilities exceeding the threshold.
2. The ECG monitor of claim 1 , further comprising a single-lead ECG configured to generate the sequence of ECG values.
3. The ECG monitor of claim 1 , the inception module comprising a concatenator configured to concatenate the plurality of feature maps to form a channel array.
4. The ECG monitor of claim 1 , wherein:
the inception module further includes a window function configured to process the sequence of ECG values into a spectrogram; and
the plurality of feature maps includes at least part of the spectrogram.
5. The ECG monitor of claim 1 , the residual neural network including a fully pre-activated downsampling residual unit having:
a downsampling CNN configured to process the plurality of feature maps with the downsampling stride;
a shortcut pooling layer configured to bypass the downsampling CNN and to pool the plurality of feature maps by the downsampling stride; and
a downsampling adder configured to add an output of the downsampling CNN and an output of the shortcut pooling layer.
6. The ECG monitor of claim 5 , the residual neural network further including, prior to the fully pre-activated downsampling residual unit, a fully pre-activated non-downsampling residual unit having:
a non-downsampling CNN configured to process the plurality of feature maps with a stride of one;
a shortcut connection configured to bypass the non-downsampling CNN; and
a non-downsampling adder configured to add an output of the non-downsampling CNN and an output of the shortcut connection.
7. The ECG monitor of claim 1 , the LSTM neural network including a sequence of memory cells.
8. The ECG monitor of claim 7 , wherein:
each memory cell of the sequence of memory cells has an input cell state and an output cell state, the output cell state being connected to the input cell state of a next memory cell of the sequence of memory cells; and
the machine-readable instructions that, when executed by the processor, control the ECG monitor to calculate the sequence of cortical-arousal probabilities include machine-readable instructions that, when executed by the processor, control the ECG monitor to:
feed the plurality of downsampled feature maps into the input cell state of a first memory cell of the sequence of memory cells; and
calculate the sequence of cortical-arousal probabilities based on the output cell state of a last memory cell of the sequence of memory cells.
9. The ECG monitor of claim 7 , the LSTM neural network including:
a fully-connected layer connected to an output of the sequence of memory cells; and
a nonlinear layer connected to an output of the fully-connected layer.
10. An electrocardiography method, comprising:
filtering, with an inception module of a deep neural network, a sequence of ECG values such that a plurality of convolutional neural networks (CNNs) of the inception module process the sequence of ECG values in parallel, each of the plurality of CNNs outputting a respective one of a plurality of feature maps;
iteratively downsampling, with a residual neural network of the deep neural network and via a feedback path that connects an output of the residual neural network to an input of the residual neural network, the plurality of feature maps to generate a respective plurality of downsampled feature maps, the residual neural network having a downsampling stride that is greater than one;
calculating a sequence of cortical-arousal probabilities by feeding the plurality of downsampled feature maps into a long short-term memory (LSTM) neural network of the deep neural network;
comparing each of the sequence of cortical-arousal probabilities to a threshold; and
displaying, on an electronic display, an indication in response to one or more of the sequence of cortical-arousal probabilities exceeding the threshold.
11. The ECG method of claim 10 , further comprising sensing the sequence of ECG values with a single-lead ECG sensor.
12. The ECG method of claim 10 , wherein said filtering comprises concatenating the plurality of feature maps to create a channel array.
13. The ECG method of claim 10 , wherein said filtering further includes applying a window function of the inception module to the sequence of ECG values to create a spectrogram, the plurality of feature maps including at least part of the spectrogram.
14. The ECG method of claim 10 , wherein said iteratively downsampling includes iteratively downsampling the plurality of feature maps with a fully pre-activated downsampling residual unit of the residual neural network, the fully pre-activated downsampling residual unit having:
a downsampling CNN that processes the plurality of feature maps with the downsampling stride;
a shortcut pooling layer that bypasses the downsampling CNN and pools the plurality of feature maps by the downsampling stride; and
a downsampling adder that adds an output of the downsampling CNN and an output of the shortcut pooling layer.
15. The ECG method of claim 14 , wherein said iteratively downsampling further includes iteratively downsampling the plurality of feature maps with a fully pre-activated non-downsampling residual unit of the residual neural network, the fully pre-activated non-downsampling residual unit having:
a non-downsampling CNN that processes the plurality of feature maps with a stride of one;
a shortcut connection that bypasses the non-downsampling CNN; and
a non-downsampling adder that adds an output of the non-downsampling CNN and an output of the shortcut connection.
16. The ECG method of claim 10 , wherein said calculating includes calculating the sequence of cortical-arousal probabilities with a sequence of memory cells of the LSTM neural network.
17. The ECG method of claim 16 , wherein said calculating the sequence of cortical-arousal probabilities with the sequence of memory cells includes:
feeding the plurality of downsampled feature maps into an input cell state of a first memory cell of the sequence of memory cells;
passing an output cell state of each memory cell of the sequence of memory cells to an input cell state of a next memory cell of the sequence of memory cells; and
calculating the sequence of cortical-arousal probabilities based on an output cell state of a last memory cell of the sequence of memory cells.
18. The ECG method of claim 16 , wherein said calculating further includes:
passing an output of the sequence of memory cells to an input of a fully-connected layer; and
passing an output of the fully-connected layer to an input of a nonlinear layer.