Electrocardiogram-based diagnosis model learning device, electrocardiogram-based diagnosis model learning method, and storage medium

1 . An electrocardiogram-based diagnosis model learning device comprising:

at least one memory configured to store instructions; and

at least one processor configured to execute the instructions to:

extract, based on data of a continuous waveform of overall lead data, an attention interval of electrocardiogram data by using a second model constituted by a convolutional neural network,

wherein the second model, in a case where a diagnosis result is outputted based on the electrocardiogram data, generates the attention interval as an output of an intermediate layer, and

wherein the attention interval is a section in which coefficients are equal to or greater than a predetermined value, and is output by an attention mechanism that is added to the second model so that the output in front of full connection is inputted to the attention mechanism;

acquire lead data for validation to calculate an accuracy index from training data, the lead data including P waves, Q waves, R waves, S waves, T waves, and U waves of an electrocardiogram waveform;

acquire a predetermined partial waveform of electrocardiogram data regarding an electrocardiogram of a subject, the predetermined partial waveform being predetermined to be specific to a target disease;

extract the lead data corresponding to the predetermined partial waveform from the lead data;

in a case where the diagnosis result output by the second model for input lead data matches ground-truth label of the input lead data, perform machine learning of a third machine learning model configured to diagnose the target disease, based on the predetermined partial waveform and a regular partial waveform, each corresponding to the input lead data, which is a partial waveform where the attention interval that does not overlap with the predetermined partial waveform appears regularly,

wherein the third machine learning model is configured to output the diagnosis result of the target disease when the electrocardiogram data corresponding to the predetermined partial waveform and the regular partial waveform is inputted to the third machine learning model; and

calculate, based on electrocardiogram validation data which is different from the electrocardiogram data, an evaluation index regarding an accuracy of the third machine learning model, wherein the evaluation index relates to an accuracy of a first model that is a model configured to output the diagnosis result of the target disease when the electrocardiogram data corresponding to the predetermined partial waveform is inputted thereto.

2 . The electrocardiogram-based diagnosis model learning device according to claim 1 ,

wherein the at least one processor is configured to execute the instructions to acquire the attention interval, based on the second model configured to output the diagnosis result of the target disease and the attention interval when the electrocardiogram data corresponding to a sequential waveform is inputted to the second model.

3 . The electrocardiogram-based diagnosis model learning device according to claim 2 ,

wherein the at least one processor is configured to execute the instructions to perform the machine learning of the third machine learning model, based on attention interval outputted by the second model when the diagnosis result of the target disease is correct.

4 . The electrocardiogram-based diagnosis model learning device according to claim 1 ,

wherein the at least one processor is configured to further execute the instructions to select a model to be used for the target disease based on the evaluation index regarding an accuracy of the third machine learning model and the evaluation index regarding an accuracy of the first model.

5 . The electrocardiogram-based diagnosis model learning device according to claim 4 ,

wherein the at least one processor is configured to further execute the instructions to:

acquire the electrocardiogram data regarding a target person to be diagnosed;

if the selected model is the first model,

generate the predetermined partial waveform based on the acquired electrocardiogram data regarding the target person, and

input the predetermined partial waveform to the first model thereby to output the diagnosis result of the target disease; and

if the selected model is the third machine learning model,

generate the predetermined partial waveform and the regular partial waveform based on the acquired electrocardiogram data regarding the target person, and

input the predetermined partial waveform and the regular partial waveform to the third machine learning model thereby to output the diagnosis result of the target disease.

6 . The electrocardiogram-based diagnosis model learning device according to claim 5 ,

wherein the at least one processor is configured to execute the instructions to display the diagnosis result on a display device to assist a healthcare worker to perform a decision making.

7 . An electrocardiogram-based diagnosis model learning method executed by a computer, comprising:

extracting, based on data of a continuous waveform of overall lead data, an attention interval of electrocardiogram data by using a second model constituted by a convolutional neural network,

wherein the second model, in a case where a diagnosis result is outputted based on the electrocardiogram data, generates the attention interval as an output of an intermediate layer, and

wherein the attention interval is a section in which coefficients are equal to or greater than a predetermined value, and is output by an attention mechanism that is added to the second model so that the output in front of full connection is inputted to the attention mechanism;

acquiring lead data for validation to calculate an accuracy index from training data, the lead data including P waves, Q waves, R waves, S waves, T waves, and U waves of an electrocardiogram waveform;

acquiring a predetermined partial waveform of electrocardiogram data regarding an electrocardiogram of a subject, the predetermined partial waveform being predetermined to be specific to a target disease;

extracting the lead data corresponding to the predetermined partial waveform from the lead data;

in a case where the diagnosis result output by the second model for input lead data matches ground-truth label of the input lead data, performing machine learning of a third machine learning model configured to diagnose the target disease, based on the predetermined partial waveform and a regular partial waveform, each corresponding to the input lead data, which is a partial waveform where the attention interval that does not overlap with the predetermined partial waveform appears regularly, the third machine learning model outputting the diagnosis result of the target disease when the electrocardiogram data corresponding to the predetermined partial waveform and the regular partial waveform is inputted to the third machine learning model; and

calculating, based on electrocardiogram validation data which is different from the electrocardiogram data, an evaluation index regarding an accuracy of the third machine learning model, wherein the evaluation index relates to an accuracy of a first model that is a model configured to output the diagnosis result of the target disease when the electrocardiogram data corresponding to the predetermined partial waveform is inputted thereto.

8 . A non-transitory computer readable storage medium storing a program executed by a computer, the program causing the computer to:

extract, based on data of a continuous waveform of overall lead data, an attention interval of electrocardiogram data by using a second model constituted by a convolutional neural network,

wherein the second model, in a case where a diagnosis result is outputted based on the electrocardiogram data, generates the attention interval as an output of an intermediate layer, and

wherein the attention interval is a section in which coefficients are equal to or greater than a predetermined value, and is output by an attention mechanism that is added to the second model so that the output in front of full connection is inputted to the attention mechanism;

acquire lead data for validation to calculate an accuracy index from training data, the lead data including P waves, Q waves, R waves, S waves, T waves, and U waves of an electrocardiogram waveform;

acquire a predetermined partial waveform of electrocardiogram data regarding an electrocardiogram of a subject, the predetermined partial waveform being predetermined to be specific to a target disease;

extract the lead data corresponding to the predetermined partial waveform from the lead data;

in a case where the diagnosis result output by the second model for input lead data matches ground-truth label of the input lead data, perform machine learning of a third machine learning model configured to diagnose the target disease, based on the predetermined partial waveform and a regular partial waveform, each corresponding to the input lead data, which is a partial waveform where the attention interval that does not overlap with the predetermined partial waveform appears regularly,

the third machine learning model being configured to output the diagnosis result of the target disease when the electrocardiogram data corresponding to the predetermined partial waveform and the regular partial waveform is inputted to the third machine learning model; and

calculate, based on electrocardiogram validation data which is different from the electrocardiogram data, an evaluation index regarding an accuracy of the third machine learning model, wherein the evaluation index relates to an accuracy of a first model that is a model configured to output the diagnosis result of the target disease when the electrocardiogram data corresponding to the predetermined partial waveform is inputted thereto.