Deep neural network pre-training method for classifying electrocardiogram (ECG) data
A deep neural network pre-training method for classifying electrocardiogram (ECG) data and a device for the same are disclosed. A method for training an ECG feature extraction model may include receiving a ECG signal, extracting one or more first features related to the ECG signal by inputting the ECG signal to a rule-based feature extractor or a neural network model, extracting at least one second feature corresponding to the at least one first feature by inputting the ECG signal to an encoder, and pre-training the ECG feature extraction model by inputting the at least one second feature into at least one of a regression function and a classification function to calculate at least one output value. The pre-training of the ECG feature extraction model may include training the encoder to minimize a loss function that is determined based on the at least one output value and the at least one first feature.
1 . A method for learning an electrocardiogram (ECG) label output model by a computing device comprising:
receiving at least one unlabeled ECG signal and at least one labeled ECG signal;
performing self-supervised learning of an ECG feature extraction model including an encoder and a rule-based feature extractor based on the at least one unlabeled ECG signal; and
performing supervised learning of the ECG label output model based on the at least one labeled ECG signal,
wherein the ECG label output model is configured to include the encoder pre-trained through the self-supervised learning of the ECG feature extraction model and a classifier configured to classify a diagnostic label;
wherein the self-supervised learning of the ECG feature extraction model includes:
extracting at least one first feature from the at least one unlabeled ECG signal using the rule-based feature extractor;
extracting at least one second feature from the at least one unlabeled ECG signal using the encoder;
obtaining at least one output value by mapping the at least one second feature to at least one of a regression function and a classification function; and
performing the self-supervised learning of the ECG feature extraction model using a loss function based on the at least one output value and the at least one first feature.
2 . The method according to claim 1 , wherein the extracting the at least one first feature includes:
classifying waveforms of an ECG signal included in the at least one unlabeled ECG signal; and
extracting the at least one first feature from the waveforms of the ECG signal according to a predetermined rule.
3 . The method according to claim 2 , wherein the classifying the waveforms of the ECG signal includes classifying the ECG signal into P waves, QRS complex, and T waves, and
wherein the extracting the at least one first feature includes extracting the at least one first feature corresponding to at least one of a heart rate, a PR interval, a QRS duration, a QT interval, a QT corrected, an R axis, a P axis, and a T axis, based on the P waves, the QRS complex, and the T waves.
4 . The method according to claim 1 , wherein the supervised learning of the ECG label output model includes:
adjusting parameters related to the encoder to minimize a difference between at least one feature extracted from the labeled ECG data and the at least one output value.
5 . The method according to claim 1 , wherein the self-supervised learning of the ECG feature extraction model further includes:
determining a parameter of at least one of the regression function and the classification function and a parameter of the encoder.
6 . The method according to claim 1 , wherein at least one output value includes:
at least one first output value calculated by inputting the at least one second feature to the regression function, and
at least one second output value calculated by inputting the at least one second feature to the classification function.
7 . The method according to claim 6 , wherein the self-supervised learning of the ECG feature extraction model further includes:
determining a first loss function based on the at least one first output value and the at least one first feature;
determining a second loss function based on the at least one second output value and the at least one first feature; and
pre-training the encoder to minimize the loss function that is determined by a combination of the first loss function and the second loss function.
8 . A computing device comprising:
a processor provided with one or more cores;
a communication interface; and
a memory,
wherein the processor is configured to:
receive at least one unlabeled ECG signal and at least one labeled ECG signal;
perform self-supervised learning of an ECG feature extraction model including an encoder and a rule-based feature extractor based on the at least one unlabeled ECG signal; and
perform supervised learning of the ECG label output model based on the at least one labeled ECG signal,
wherein the ECG label output model is configured to include the encoder pre-trained through the self-supervised learning of the ECG feature extraction model and a classifier configured to classify a diagnostic label,
wherein the self-supervised learning of the ECG feature extraction model includes extracting at least one first feature from the at least one unlabeled ECG signal using the rule-based feature extractor, extracting at least one second feature form the at least one unlabeled ECG signal using the encoder of the ECG label output model, obtaining at least one output value by mapping the at least one second feature to at least one of a regression function and a classification function for an ECG feature extraction model, and performing the self-supervised learning of the ECG feature extraction model using a loss function based on the at least one output value and the at least one first feature.
9 . The computing device of claim 8 , wherein the processor is further configured to classify waveforms of an ECG signal included in the at least one unlabeled ECG data, and extract the at least one first feature from the waveforms of the ECG signal according to a predetermined rule.
10 . The computing device of claim 9 , wherein the processor is further configured to:
classify the ECG signal into P waves, QRS complex, and T waves, and
extract the at least one first feature corresponding to at least one of a heart rate, a PR interval, a QRS duration, a QT interval, a QT corrected, an R axis, a P axis, and a T axis, based on the P waves, the QRS complex, and the T waves.
11 . The computing device of claim 8 , wherein, in the self-supervised learning of the ECG feature extraction model, the processor is further configured to determine a parameter of at least one of the regression function and the classification function and a parameter of the encoder.
12 . The computing device of claim 8 , wherein the self-supervised learning of the ECG feature extraction model includes:
at least one first output value calculated by inputting the at least one second feature to the regression function, and
at least one second output value calculated by inputting the at least one second feature to the classification function.
13 . The computing device of claim 12 , wherein, in the self-supervised learning of the ECG feature extraction model, the processor is further configured to:
determine a first loss function based on the at least one first output value and the at least one first feature;
determine a second loss function based on the at least one second output value and the at least one first feature; and
pre-train the encoder to minimize the loss function that is determined by a combination of the first loss function and the second loss function.