Medical diagnostic tool with neural model trained through machine learning for predicting coronary disease from ECG signals
A diagnostic tool includes a sensor for capturing at least one biosignal produced by a patient's heart and a computer device that implements a neural network iteratively trained via machine learning to generate a prediction about a heart condition of the patient. After the neural network is trained, the computer device can convert the at least one biosignal to a multi-dimensional input matrix for the deep neural network generated from a number (N) of biosignals captured by the sensor. The computer device then processes the multi-dimensional input matrix through the deep neural network, which subsequently outputs the prediction about the heart condition of the patient.
1 . A diagnostic tool comprising:
a sensor for capturing N biosignals produced by a patient's heart, where N is greater than or equal to 1; and
a computer device that implements a deep neural network that is trained iteratively through machine learning to generate multiple heart-condition-diagnostic outputs, wherein each of the multiple heart-condition-diagnostic outputs is indicative of a heart condition of the patient, and wherein, after the deep neural network is trained to generate the multiple heart-condition-diagnostic outputs, the computer device is configured to:
convert, with an autoencoder, the N biosignals to a multi-dimensional input matrix for the deep neural network generated from the N biosignals captured by the sensor, wherein the multi-dimensional matrix comprises a batch axis, a signal axis, and a lead axis, generated from the N biosignals, and wherein each of the N biosignals comprises a time component that is at least “T” seconds in duration, where T is at least one second; and
process the multi-dimensional input matrix through the deep neural network to generate the multiple heart-condition-diagnostic outputs,
wherein:
the multi-dimensional matrix is inputted to the deep neural network;
the deep neural network produces a feature vector from the multi-dimensional matrix; and
the multiple heart-condition-diagnostic outputs are based on the feature vector.
2 . The diagnostic tool of claim 1 , wherein the multiple heart-condition-diagnostic outputs comprise at least one prediction of coronary artery disease (CAD) of the patient.
3 . The diagnostic tool of claim 2 , wherein the multiple heart-condition-diagnostic outputs further comprise at least one additional prediction selected from the group consisting of:
a Myocardial Infarction (MI), a risk of a Major Adverse Cardiovascular Event (MACE), a left anterior descending (LAD) coronary artery fractional flow reserve (FFR), an atherosclerotic cardiovascular disease (ACVD), cardiac hypertrophy, ventricle morphology, an abnormal ST-T wave, a conduction disorder, and a “70%” disease threshold.
4 . The diagnostic tool of claim 1 , wherein:
the deep neural network comprises a dense neural network (DenseNet), wherein each layer of the DenseNet after an input layer receives inputs from all preceding layers in the DenseNet;
the DenseNet produces a feature vector from the multi-dimensional matrix;
the deep neural network further comprises a classifier whose outputs are the multiple heart-condition-diagnostic outputs for the patient; and
the input to the classifier comprises a concatenation of the feature vector from the DenseNet and a latent space representation from the autoencoder.
5 . The diagnostic tool of claim 1 , wherein the deep neural network comprises a residual network, wherein the residual network comprises at least one residual connection.
6 . The diagnostic tool of claim 1 , wherein the deep neural network comprises a vision transformer, wherein the vision transformer comprises a transformer encoder.
7 . The diagnostic tool of claim 1 , wherein:
the N biosignals comprise N electrocardiogram (ECG) signals;
N is less than 12; and
the autoencoder is further trained via machine learning to generate (15-N) generated biosignals based, at least in part, on a reconstruction loss function.
8 . The diagnostic tool of claim 7 , wherein the N ECG signals plus the (15-N) generated biosignals collectively comprise 12 ECG signals and 3 Frank lead signals.
9 . The diagnostic tool of claim 7 , wherein the reconstruction loss function calculates a squared difference between a fast Fourier transform (FFT) between an input of the sensor and an output of the autoencoder.
10 . The diagnostic tool of claim 1 , wherein the multiple heart-condition-diagnostic outputs comprise at least one numerical value selected from the group consisting of: a coronary artery calcium score, an absolute Agatston score, a multi-vessel fractional flow reserve value, and an instantaneous wave-free ratio.
11 . The diagnostic tool of claim 1 , wherein the autoencoder is further programmed to compute a prediction of T additional seconds of up to 15 biosignals for the patient's heart.
12 . The diagnostic tool of claim 1 , wherein the autoencoder comprises:
an encoder configured to perform a lossy compression of the N biosignals captured by the sensor, wherein an output of the lossy compression is latent space representation; and
a decoder configured to receive the latent space from the encoder and convert the latent space representation into the multi-dimensional input matrix.
13 . The diagnostic tool of claim 1 , wherein the multiple heart-condition-diagnostic outputs comprise at least one of a beat classification and a rhythm classification.
14 . The diagnostic tool of claim 1 , further comprising a display coupled to the computer device, wherein the display is configured to visually represent the multiple heart-condition-diagnostic outputs of the deep neural network.
15 . A method comprising:
training, with a computer system, intertively through machine learning, a deep neural network to generate multiple heart-condition-diagnostic outputs, wherein each of the multiple heart-condition-diagnostic outputs is indicative of a heart condition of a patient and wherein the deep neural network comprises a plurality of layers; and
after training the deep neural network to generate the multiple heart-condition-diagnostic outputs:
capturing, by a sensor, N biosignals produced by a patient's heart, where N is greater than or equal to 1; and
converting, with an autoencoder of the computer system, the N biosignals to a multi-dimensional input matrix for the deep neural network generated from the N biosignals captured by the sensor, wherein the multi-dimensional matrix comprises a batch axis, a signal axis, and a lead axis, generated from the N biosignals, and wherein each of the N biosignals comprises a time component that is at least “T” seconds in duration, where T is at least one second; and
processing the multi-dimensional input matrix through the deep neural network to generate the multiple heart-condition-diagnostic outputs, wherein the deep neural network produces a feature vector from the multi-dimensional matrix and the multiple heart-condition-diagnostic outputs are based on the feature vector.
16 . The method of claim 15 , wherein:
the N biosignals comprise N electrocardiogram (ECG) signals;
N is less than 12; and
the method further comprises training, by the computer system, the autoencoder via machine learning to generate (15-N) generated biosignals based, at least in part, on a reconstruction loss function.
17 . The method of claim 15 , wherein the N ECG signals plus the (15-N) generated biosignals collectively comprise 12 ECG signals and 3 Frank lead signals.
18 . The method of claim 15 , wherein the deep neural network comprises a residual network, wherein the residual network comprises at least one residual connection.
19 . The method of claim 15 , wherein the deep neural network comprises a vision transformer, wherein the vision transformer comprises a transformer encoder.
20 . The method of claim 15 , wherein the autoencoder comprises:
an encoder configured to perform a lossy compression of the N biosignals captured by the sensor, wherein an output of the lossy compression is latent space representation; and
a decoder configured to receive the latent space from the encoder and convert the latent space representation into the multi-dimensional input matrix.